5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciUtilities.hpp"
26 #include "classfile/javaClasses.hpp"
27 #include "ci/ciObjArray.hpp"
28 #include "asm/register.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subtypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "utilities/bitMap.inline.hpp"
50 #include "utilities/powerOfTwo.hpp"
51 #include "utilities/growableArray.hpp"
52
53 //----------------------------GraphKit-----------------------------------------
54 // Main utility constructor.
55 GraphKit::GraphKit(JVMState* jvms)
56 : Phase(Phase::Parser),
57 _env(C->env()),
58 _gvn(*C->initial_gvn()),
59 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
60 {
61 _exceptions = jvms->map()->next_exception();
62 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
63 set_jvms(jvms);
64 }
65
66 // Private constructor for parser.
67 GraphKit::GraphKit()
68 : Phase(Phase::Parser),
69 _env(C->env()),
70 _gvn(*C->initial_gvn()),
71 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
72 {
73 _exceptions = nullptr;
74 set_map(nullptr);
75 debug_only(_sp = -99);
76 debug_only(set_bci(-99));
77 }
78
79
80
81 //---------------------------clean_stack---------------------------------------
82 // Clear away rubbish from the stack area of the JVM state.
83 // This destroys any arguments that may be waiting on the stack.
328 }
329 static inline void add_one_req(Node* dstphi, Node* src) {
330 assert(is_hidden_merge(dstphi), "must be a special merge node");
331 assert(!is_hidden_merge(src), "must not be a special merge node");
332 dstphi->add_req(src);
333 }
334
335 //-----------------------combine_exception_states------------------------------
336 // This helper function combines exception states by building phis on a
337 // specially marked state-merging region. These regions and phis are
338 // untransformed, and can build up gradually. The region is marked by
339 // having a control input of its exception map, rather than null. Such
340 // regions do not appear except in this function, and in use_exception_state.
341 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
342 if (failing_internal()) {
343 return; // dying anyway...
344 }
345 JVMState* ex_jvms = ex_map->_jvms;
346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
351 assert(ex_map->req() == phi_map->req(), "matching maps");
352 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
353 Node* hidden_merge_mark = root();
354 Node* region = phi_map->control();
355 MergeMemNode* phi_mem = phi_map->merged_memory();
356 MergeMemNode* ex_mem = ex_map->merged_memory();
357 if (region->in(0) != hidden_merge_mark) {
358 // The control input is not (yet) a specially-marked region in phi_map.
359 // Make it so, and build some phis.
360 region = new RegionNode(2);
361 _gvn.set_type(region, Type::CONTROL);
362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
363 region->init_req(1, phi_map->control());
364 phi_map->set_control(region);
365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
366 record_for_igvn(io_phi);
367 _gvn.set_type(io_phi, Type::ABIO);
368 phi_map->set_i_o(io_phi);
856 if (PrintMiscellaneous && (Verbose || WizardMode)) {
857 tty->print_cr("Zombie local %d: ", local);
858 jvms->dump();
859 }
860 return false;
861 }
862 }
863 }
864 return true;
865 }
866
867 #endif //ASSERT
868
869 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
870 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
871 ciMethod* cur_method = jvms->method();
872 int cur_bci = jvms->bci();
873 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
874 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
875 return Interpreter::bytecode_should_reexecute(code) ||
876 (is_anewarray && code == Bytecodes::_multianewarray);
877 // Reexecute _multianewarray bytecode which was replaced with
878 // sequence of [a]newarray. See Parse::do_multianewarray().
879 //
880 // Note: interpreter should not have it set since this optimization
881 // is limited by dimensions and guarded by flag so in some cases
882 // multianewarray() runtime calls will be generated and
883 // the bytecode should not be reexecutes (stack will not be reset).
884 } else {
885 return false;
886 }
887 }
888
889 // Helper function for adding JVMState and debug information to node
890 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
891 // Add the safepoint edges to the call (or other safepoint).
892
893 // Make sure dead locals are set to top. This
894 // should help register allocation time and cut down on the size
895 // of the deoptimization information.
896 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
947 }
948
949 // Presize the call:
950 DEBUG_ONLY(uint non_debug_edges = call->req());
951 call->add_req_batch(top(), youngest_jvms->debug_depth());
952 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
953
954 // Set up edges so that the call looks like this:
955 // Call [state:] ctl io mem fptr retadr
956 // [parms:] parm0 ... parmN
957 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
958 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
959 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
960 // Note that caller debug info precedes callee debug info.
961
962 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
963 uint debug_ptr = call->req();
964
965 // Loop over the map input edges associated with jvms, add them
966 // to the call node, & reset all offsets to match call node array.
967 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
968 uint debug_end = debug_ptr;
969 uint debug_start = debug_ptr - in_jvms->debug_size();
970 debug_ptr = debug_start; // back up the ptr
971
972 uint p = debug_start; // walks forward in [debug_start, debug_end)
973 uint j, k, l;
974 SafePointNode* in_map = in_jvms->map();
975 out_jvms->set_map(call);
976
977 if (can_prune_locals) {
978 assert(in_jvms->method() == out_jvms->method(), "sanity");
979 // If the current throw can reach an exception handler in this JVMS,
980 // then we must keep everything live that can reach that handler.
981 // As a quick and dirty approximation, we look for any handlers at all.
982 if (in_jvms->method()->has_exception_handlers()) {
983 can_prune_locals = false;
984 }
985 }
986
987 // Add the Locals
988 k = in_jvms->locoff();
989 l = in_jvms->loc_size();
990 out_jvms->set_locoff(p);
991 if (!can_prune_locals) {
992 for (j = 0; j < l; j++)
993 call->set_req(p++, in_map->in(k+j));
994 } else {
995 p += l; // already set to top above by add_req_batch
996 }
997
998 // Add the Expression Stack
999 k = in_jvms->stkoff();
1000 l = in_jvms->sp();
1001 out_jvms->set_stkoff(p);
1002 if (!can_prune_locals) {
1003 for (j = 0; j < l; j++)
1004 call->set_req(p++, in_map->in(k+j));
1005 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1006 // Divide stack into {S0,...,S1}, where S0 is set to top.
1007 uint s1 = stack_slots_not_pruned;
1008 stack_slots_not_pruned = 0; // for next iteration
1009 if (s1 > l) s1 = l;
1010 uint s0 = l - s1;
1011 p += s0; // skip the tops preinstalled by add_req_batch
1012 for (j = s0; j < l; j++)
1013 call->set_req(p++, in_map->in(k+j));
1014 } else {
1015 p += l; // already set to top above by add_req_batch
1016 }
1017
1018 // Add the Monitors
1019 k = in_jvms->monoff();
1020 l = in_jvms->mon_size();
1021 out_jvms->set_monoff(p);
1022 for (j = 0; j < l; j++)
1023 call->set_req(p++, in_map->in(k+j));
1024
1025 // Copy any scalar object fields.
1026 k = in_jvms->scloff();
1027 l = in_jvms->scl_size();
1028 out_jvms->set_scloff(p);
1029 for (j = 0; j < l; j++)
1030 call->set_req(p++, in_map->in(k+j));
1031
1032 // Finish the new jvms.
1033 out_jvms->set_endoff(p);
1034
1035 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1036 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1037 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1038 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1039 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1040 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1041
1042 // Update the two tail pointers in parallel.
1043 out_jvms = out_jvms->caller();
1044 in_jvms = in_jvms->caller();
1045 }
1046
1047 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1048
1049 // Test the correctness of JVMState::debug_xxx accessors:
1050 assert(call->jvms()->debug_start() == non_debug_edges, "");
1051 assert(call->jvms()->debug_end() == call->req(), "");
1052 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1053 }
1054
1055 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1056 Bytecodes::Code code = java_bc();
1057 if (code == Bytecodes::_wide) {
1058 code = method()->java_code_at_bci(bci() + 1);
1059 }
1060
1061 if (code != Bytecodes::_illegal) {
1062 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1193 Node* GraphKit::ConvI2UL(Node* offset) {
1194 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1195 if (offset_con != (juint) Type::OffsetBot) {
1196 return longcon((julong) offset_con);
1197 }
1198 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1199 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1200 return _gvn.transform( new AndLNode(conv, mask) );
1201 }
1202
1203 Node* GraphKit::ConvL2I(Node* offset) {
1204 // short-circuit a common case
1205 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1206 if (offset_con != (jlong)Type::OffsetBot) {
1207 return intcon((int) offset_con);
1208 }
1209 return _gvn.transform( new ConvL2INode(offset));
1210 }
1211
1212 //-------------------------load_object_klass-----------------------------------
1213 Node* GraphKit::load_object_klass(Node* obj) {
1214 // Special-case a fresh allocation to avoid building nodes:
1215 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1216 if (akls != nullptr) return akls;
1217 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1218 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1219 }
1220
1221 //-------------------------load_array_length-----------------------------------
1222 Node* GraphKit::load_array_length(Node* array) {
1223 // Special-case a fresh allocation to avoid building nodes:
1224 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1225 Node *alen;
1226 if (alloc == nullptr) {
1227 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1228 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1229 } else {
1230 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1231 }
1232 return alen;
1233 }
1234
1235 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1236 const TypeOopPtr* oop_type,
1237 bool replace_length_in_map) {
1238 Node* length = alloc->Ideal_length();
1247 replace_in_map(length, ccast);
1248 }
1249 return ccast;
1250 }
1251 }
1252 return length;
1253 }
1254
1255 //------------------------------do_null_check----------------------------------
1256 // Helper function to do a null pointer check. Returned value is
1257 // the incoming address with null casted away. You are allowed to use the
1258 // not-null value only if you are control dependent on the test.
1259 #ifndef PRODUCT
1260 extern uint explicit_null_checks_inserted,
1261 explicit_null_checks_elided;
1262 #endif
1263 Node* GraphKit::null_check_common(Node* value, BasicType type,
1264 // optional arguments for variations:
1265 bool assert_null,
1266 Node* *null_control,
1267 bool speculative) {
1268 assert(!assert_null || null_control == nullptr, "not both at once");
1269 if (stopped()) return top();
1270 NOT_PRODUCT(explicit_null_checks_inserted++);
1271
1272 // Construct null check
1273 Node *chk = nullptr;
1274 switch(type) {
1275 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1276 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1277 case T_ARRAY : // fall through
1278 type = T_OBJECT; // simplify further tests
1279 case T_OBJECT : {
1280 const Type *t = _gvn.type( value );
1281
1282 const TypeOopPtr* tp = t->isa_oopptr();
1283 if (tp != nullptr && !tp->is_loaded()
1284 // Only for do_null_check, not any of its siblings:
1285 && !assert_null && null_control == nullptr) {
1286 // Usually, any field access or invocation on an unloaded oop type
1287 // will simply fail to link, since the statically linked class is
1288 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1289 // the static class is loaded but the sharper oop type is not.
1290 // Rather than checking for this obscure case in lots of places,
1291 // we simply observe that a null check on an unloaded class
1355 }
1356 Node *oldcontrol = control();
1357 set_control(cfg);
1358 Node *res = cast_not_null(value);
1359 set_control(oldcontrol);
1360 NOT_PRODUCT(explicit_null_checks_elided++);
1361 return res;
1362 }
1363 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1364 if (cfg == nullptr) break; // Quit at region nodes
1365 depth++;
1366 }
1367 }
1368
1369 //-----------
1370 // Branch to failure if null
1371 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1372 Deoptimization::DeoptReason reason;
1373 if (assert_null) {
1374 reason = Deoptimization::reason_null_assert(speculative);
1375 } else if (type == T_OBJECT) {
1376 reason = Deoptimization::reason_null_check(speculative);
1377 } else {
1378 reason = Deoptimization::Reason_div0_check;
1379 }
1380 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1381 // ciMethodData::has_trap_at will return a conservative -1 if any
1382 // must-be-null assertion has failed. This could cause performance
1383 // problems for a method after its first do_null_assert failure.
1384 // Consider using 'Reason_class_check' instead?
1385
1386 // To cause an implicit null check, we set the not-null probability
1387 // to the maximum (PROB_MAX). For an explicit check the probability
1388 // is set to a smaller value.
1389 if (null_control != nullptr || too_many_traps(reason)) {
1390 // probability is less likely
1391 ok_prob = PROB_LIKELY_MAG(3);
1392 } else if (!assert_null &&
1393 (ImplicitNullCheckThreshold > 0) &&
1394 method() != nullptr &&
1395 (method()->method_data()->trap_count(reason)
1429 }
1430
1431 if (assert_null) {
1432 // Cast obj to null on this path.
1433 replace_in_map(value, zerocon(type));
1434 return zerocon(type);
1435 }
1436
1437 // Cast obj to not-null on this path, if there is no null_control.
1438 // (If there is a null_control, a non-null value may come back to haunt us.)
1439 if (type == T_OBJECT) {
1440 Node* cast = cast_not_null(value, false);
1441 if (null_control == nullptr || (*null_control) == top())
1442 replace_in_map(value, cast);
1443 value = cast;
1444 }
1445
1446 return value;
1447 }
1448
1449
1450 //------------------------------cast_not_null----------------------------------
1451 // Cast obj to not-null on this path
1452 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1453 const Type *t = _gvn.type(obj);
1454 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1455 // Object is already not-null?
1456 if( t == t_not_null ) return obj;
1457
1458 Node* cast = new CastPPNode(control(), obj,t_not_null);
1459 cast = _gvn.transform( cast );
1460
1461 // Scan for instances of 'obj' in the current JVM mapping.
1462 // These instances are known to be not-null after the test.
1463 if (do_replace_in_map)
1464 replace_in_map(obj, cast);
1465
1466 return cast; // Return casted value
1467 }
1468
1469 // Sometimes in intrinsics, we implicitly know an object is not null
1470 // (there's no actual null check) so we can cast it to not null. In
1471 // the course of optimizations, the input to the cast can become null.
1472 // In that case that data path will die and we need the control path
1473 // to become dead as well to keep the graph consistent. So we have to
1474 // add a check for null for which one branch can't be taken. It uses
1475 // an OpaqueNotNull node that will cause the check to be removed after loop
1476 // opts so the test goes away and the compiled code doesn't execute a
1477 // useless check.
1478 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1479 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1480 return value;
1481 }
1482 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1483 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1484 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1485 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1486 _gvn.set_type(iff, iff->Value(&_gvn));
1487 if (!tst->is_Con()) {
1488 record_for_igvn(iff);
1561 // These are layered on top of the factory methods in LoadNode and StoreNode,
1562 // and integrate with the parser's memory state and _gvn engine.
1563 //
1564
1565 // factory methods in "int adr_idx"
1566 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1567 MemNode::MemOrd mo,
1568 LoadNode::ControlDependency control_dependency,
1569 bool require_atomic_access,
1570 bool unaligned,
1571 bool mismatched,
1572 bool unsafe,
1573 uint8_t barrier_data) {
1574 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1575 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1576 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1577 debug_only(adr_type = C->get_adr_type(adr_idx));
1578 Node* mem = memory(adr_idx);
1579 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1580 ld = _gvn.transform(ld);
1581 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1582 // Improve graph before escape analysis and boxing elimination.
1583 record_for_igvn(ld);
1584 if (ld->is_DecodeN()) {
1585 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1586 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1587 // a Phi). Recording such cases is still perfectly sound, but may be
1588 // unnecessary and result in some minor IGVN overhead.
1589 record_for_igvn(ld->in(1));
1590 }
1591 }
1592 return ld;
1593 }
1594
1595 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1596 MemNode::MemOrd mo,
1597 bool require_atomic_access,
1598 bool unaligned,
1599 bool mismatched,
1600 bool unsafe,
1614 if (unsafe) {
1615 st->as_Store()->set_unsafe_access();
1616 }
1617 st->as_Store()->set_barrier_data(barrier_data);
1618 st = _gvn.transform(st);
1619 set_memory(st, adr_idx);
1620 // Back-to-back stores can only remove intermediate store with DU info
1621 // so push on worklist for optimizer.
1622 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1623 record_for_igvn(st);
1624
1625 return st;
1626 }
1627
1628 Node* GraphKit::access_store_at(Node* obj,
1629 Node* adr,
1630 const TypePtr* adr_type,
1631 Node* val,
1632 const Type* val_type,
1633 BasicType bt,
1634 DecoratorSet decorators) {
1635 // Transformation of a value which could be null pointer (CastPP #null)
1636 // could be delayed during Parse (for example, in adjust_map_after_if()).
1637 // Execute transformation here to avoid barrier generation in such case.
1638 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1639 val = _gvn.makecon(TypePtr::NULL_PTR);
1640 }
1641
1642 if (stopped()) {
1643 return top(); // Dead path ?
1644 }
1645
1646 assert(val != nullptr, "not dead path");
1647
1648 C2AccessValuePtr addr(adr, adr_type);
1649 C2AccessValue value(val, val_type);
1650 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1651 if (access.is_raw()) {
1652 return _barrier_set->BarrierSetC2::store_at(access, value);
1653 } else {
1654 return _barrier_set->store_at(access, value);
1655 }
1656 }
1657
1658 Node* GraphKit::access_load_at(Node* obj, // containing obj
1659 Node* adr, // actual address to store val at
1660 const TypePtr* adr_type,
1661 const Type* val_type,
1662 BasicType bt,
1663 DecoratorSet decorators) {
1664 if (stopped()) {
1665 return top(); // Dead path ?
1666 }
1667
1668 C2AccessValuePtr addr(adr, adr_type);
1669 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1670 if (access.is_raw()) {
1671 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1672 } else {
1673 return _barrier_set->load_at(access, val_type);
1674 }
1675 }
1676
1677 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1678 const Type* val_type,
1679 BasicType bt,
1680 DecoratorSet decorators) {
1681 if (stopped()) {
1682 return top(); // Dead path ?
1683 }
1684
1685 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1686 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1687 if (access.is_raw()) {
1688 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1689 } else {
1754 Node* new_val,
1755 const Type* value_type,
1756 BasicType bt,
1757 DecoratorSet decorators) {
1758 C2AccessValuePtr addr(adr, adr_type);
1759 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1760 if (access.is_raw()) {
1761 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1762 } else {
1763 return _barrier_set->atomic_add_at(access, new_val, value_type);
1764 }
1765 }
1766
1767 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1768 return _barrier_set->clone(this, src, dst, size, is_array);
1769 }
1770
1771 //-------------------------array_element_address-------------------------
1772 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1773 const TypeInt* sizetype, Node* ctrl) {
1774 uint shift = exact_log2(type2aelembytes(elembt));
1775 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1776
1777 // short-circuit a common case (saves lots of confusing waste motion)
1778 jint idx_con = find_int_con(idx, -1);
1779 if (idx_con >= 0) {
1780 intptr_t offset = header + ((intptr_t)idx_con << shift);
1781 return basic_plus_adr(ary, offset);
1782 }
1783
1784 // must be correct type for alignment purposes
1785 Node* base = basic_plus_adr(ary, header);
1786 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1787 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1788 return basic_plus_adr(ary, base, scale);
1789 }
1790
1791 //-------------------------load_array_element-------------------------
1792 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1793 const Type* elemtype = arytype->elem();
1794 BasicType elembt = elemtype->array_element_basic_type();
1795 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1796 if (elembt == T_NARROWOOP) {
1797 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1798 }
1799 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1800 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1801 return ld;
1802 }
1803
1804 //-------------------------set_arguments_for_java_call-------------------------
1805 // Arguments (pre-popped from the stack) are taken from the JVMS.
1806 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1807 // Add the call arguments:
1808 uint nargs = call->method()->arg_size();
1809 for (uint i = 0; i < nargs; i++) {
1810 Node* arg = argument(i);
1811 call->init_req(i + TypeFunc::Parms, arg);
1812 }
1813 }
1814
1815 //---------------------------set_edges_for_java_call---------------------------
1816 // Connect a newly created call into the current JVMS.
1817 // A return value node (if any) is returned from set_edges_for_java_call.
1818 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1819
1820 // Add the predefined inputs:
1821 call->init_req( TypeFunc::Control, control() );
1822 call->init_req( TypeFunc::I_O , i_o() );
1823 call->init_req( TypeFunc::Memory , reset_memory() );
1824 call->init_req( TypeFunc::FramePtr, frameptr() );
1825 call->init_req( TypeFunc::ReturnAdr, top() );
1826
1827 add_safepoint_edges(call, must_throw);
1828
1829 Node* xcall = _gvn.transform(call);
1830
1831 if (xcall == top()) {
1832 set_control(top());
1833 return;
1834 }
1835 assert(xcall == call, "call identity is stable");
1836
1837 // Re-use the current map to produce the result.
1838
1839 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1840 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1841 set_all_memory_call(xcall, separate_io_proj);
1842
1843 //return xcall; // no need, caller already has it
1844 }
1845
1846 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1847 if (stopped()) return top(); // maybe the call folded up?
1848
1849 // Capture the return value, if any.
1850 Node* ret;
1851 if (call->method() == nullptr ||
1852 call->method()->return_type()->basic_type() == T_VOID)
1853 ret = top();
1854 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1855
1856 // Note: Since any out-of-line call can produce an exception,
1857 // we always insert an I_O projection from the call into the result.
1858
1859 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1860
1861 if (separate_io_proj) {
1862 // The caller requested separate projections be used by the fall
1863 // through and exceptional paths, so replace the projections for
1864 // the fall through path.
1865 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1866 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1867 }
1868 return ret;
1869 }
1870
1871 //--------------------set_predefined_input_for_runtime_call--------------------
1872 // Reading and setting the memory state is way conservative here.
1873 // The real problem is that I am not doing real Type analysis on memory,
1874 // so I cannot distinguish card mark stores from other stores. Across a GC
1875 // point the Store Barrier and the card mark memory has to agree. I cannot
1876 // have a card mark store and its barrier split across the GC point from
1877 // either above or below. Here I get that to happen by reading ALL of memory.
1878 // A better answer would be to separate out card marks from other memory.
1879 // For now, return the input memory state, so that it can be reused
1880 // after the call, if this call has restricted memory effects.
1881 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1882 // Set fixed predefined input arguments
1883 Node* memory = reset_memory();
1884 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
1885 call->init_req( TypeFunc::Control, control() );
1886 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1887 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1938 if (use->is_MergeMem()) {
1939 wl.push(use);
1940 }
1941 }
1942 }
1943
1944 // Replace the call with the current state of the kit.
1945 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
1946 JVMState* ejvms = nullptr;
1947 if (has_exceptions()) {
1948 ejvms = transfer_exceptions_into_jvms();
1949 }
1950
1951 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1952 ReplacedNodes replaced_nodes_exception;
1953 Node* ex_ctl = top();
1954
1955 SafePointNode* final_state = stop();
1956
1957 // Find all the needed outputs of this call
1958 CallProjections callprojs;
1959 call->extract_projections(&callprojs, true, do_asserts);
1960
1961 Unique_Node_List wl;
1962 Node* init_mem = call->in(TypeFunc::Memory);
1963 Node* final_mem = final_state->in(TypeFunc::Memory);
1964 Node* final_ctl = final_state->in(TypeFunc::Control);
1965 Node* final_io = final_state->in(TypeFunc::I_O);
1966
1967 // Replace all the old call edges with the edges from the inlining result
1968 if (callprojs.fallthrough_catchproj != nullptr) {
1969 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1970 }
1971 if (callprojs.fallthrough_memproj != nullptr) {
1972 if (final_mem->is_MergeMem()) {
1973 // Parser's exits MergeMem was not transformed but may be optimized
1974 final_mem = _gvn.transform(final_mem);
1975 }
1976 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1977 add_mergemem_users_to_worklist(wl, final_mem);
1978 }
1979 if (callprojs.fallthrough_ioproj != nullptr) {
1980 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1981 }
1982
1983 // Replace the result with the new result if it exists and is used
1984 if (callprojs.resproj != nullptr && result != nullptr) {
1985 C->gvn_replace_by(callprojs.resproj, result);
1986 }
1987
1988 if (ejvms == nullptr) {
1989 // No exception edges to simply kill off those paths
1990 if (callprojs.catchall_catchproj != nullptr) {
1991 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1992 }
1993 if (callprojs.catchall_memproj != nullptr) {
1994 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1995 }
1996 if (callprojs.catchall_ioproj != nullptr) {
1997 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1998 }
1999 // Replace the old exception object with top
2000 if (callprojs.exobj != nullptr) {
2001 C->gvn_replace_by(callprojs.exobj, C->top());
2002 }
2003 } else {
2004 GraphKit ekit(ejvms);
2005
2006 // Load my combined exception state into the kit, with all phis transformed:
2007 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2008 replaced_nodes_exception = ex_map->replaced_nodes();
2009
2010 Node* ex_oop = ekit.use_exception_state(ex_map);
2011
2012 if (callprojs.catchall_catchproj != nullptr) {
2013 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2014 ex_ctl = ekit.control();
2015 }
2016 if (callprojs.catchall_memproj != nullptr) {
2017 Node* ex_mem = ekit.reset_memory();
2018 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2019 add_mergemem_users_to_worklist(wl, ex_mem);
2020 }
2021 if (callprojs.catchall_ioproj != nullptr) {
2022 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2023 }
2024
2025 // Replace the old exception object with the newly created one
2026 if (callprojs.exobj != nullptr) {
2027 C->gvn_replace_by(callprojs.exobj, ex_oop);
2028 }
2029 }
2030
2031 // Disconnect the call from the graph
2032 call->disconnect_inputs(C);
2033 C->gvn_replace_by(call, C->top());
2034
2035 // Clean up any MergeMems that feed other MergeMems since the
2036 // optimizer doesn't like that.
2037 while (wl.size() > 0) {
2038 _gvn.transform(wl.pop());
2039 }
2040
2041 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2042 replaced_nodes.apply(C, final_ctl);
2043 }
2044 if (!ex_ctl->is_top() && do_replaced_nodes) {
2045 replaced_nodes_exception.apply(C, ex_ctl);
2046 }
2047 }
2048
2049
2050 //------------------------------increment_counter------------------------------
2051 // for statistics: increment a VM counter by 1
2052
2053 void GraphKit::increment_counter(address counter_addr) {
2054 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2055 increment_counter(adr1);
2056 }
2057
2058 void GraphKit::increment_counter(Node* counter_addr) {
2059 Node* ctrl = control();
2060 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2061 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2221 *
2222 * @param n node that the type applies to
2223 * @param exact_kls type from profiling
2224 * @param maybe_null did profiling see null?
2225 *
2226 * @return node with improved type
2227 */
2228 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2229 const Type* current_type = _gvn.type(n);
2230 assert(UseTypeSpeculation, "type speculation must be on");
2231
2232 const TypePtr* speculative = current_type->speculative();
2233
2234 // Should the klass from the profile be recorded in the speculative type?
2235 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2236 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2237 const TypeOopPtr* xtype = tklass->as_instance_type();
2238 assert(xtype->klass_is_exact(), "Should be exact");
2239 // Any reason to believe n is not null (from this profiling or a previous one)?
2240 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2241 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2242 // record the new speculative type's depth
2243 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2244 speculative = speculative->with_inline_depth(jvms()->depth());
2245 } else if (current_type->would_improve_ptr(ptr_kind)) {
2246 // Profiling report that null was never seen so we can change the
2247 // speculative type to non null ptr.
2248 if (ptr_kind == ProfileAlwaysNull) {
2249 speculative = TypePtr::NULL_PTR;
2250 } else {
2251 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2252 const TypePtr* ptr = TypePtr::NOTNULL;
2253 if (speculative != nullptr) {
2254 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2255 } else {
2256 speculative = ptr;
2257 }
2258 }
2259 }
2260
2261 if (speculative != current_type->speculative()) {
2262 // Build a type with a speculative type (what we think we know
2263 // about the type but will need a guard when we use it)
2264 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2265 // We're changing the type, we need a new CheckCast node to carry
2266 // the new type. The new type depends on the control: what
2267 // profiling tells us is only valid from here as far as we can
2268 // tell.
2269 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2270 cast = _gvn.transform(cast);
2271 replace_in_map(n, cast);
2272 n = cast;
2273 }
2274
2275 return n;
2276 }
2277
2278 /**
2279 * Record profiling data from receiver profiling at an invoke with the
2280 * type system so that it can propagate it (speculation)
2281 *
2282 * @param n receiver node
2283 *
2284 * @return node with improved type
2285 */
2286 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2287 if (!UseTypeSpeculation) {
2288 return n;
2289 }
2290 ciKlass* exact_kls = profile_has_unique_klass();
2291 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2292 if ((java_bc() == Bytecodes::_checkcast ||
2293 java_bc() == Bytecodes::_instanceof ||
2294 java_bc() == Bytecodes::_aastore) &&
2295 method()->method_data()->is_mature()) {
2296 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2297 if (data != nullptr) {
2298 if (!data->as_BitData()->null_seen()) {
2299 ptr_kind = ProfileNeverNull;
2300 } else {
2301 assert(data->is_ReceiverTypeData(), "bad profile data type");
2302 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2303 uint i = 0;
2304 for (; i < call->row_limit(); i++) {
2305 ciKlass* receiver = call->receiver(i);
2306 if (receiver != nullptr) {
2307 break;
2308 }
2309 }
2310 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2311 }
2312 }
2313 }
2314 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2315 }
2316
2317 /**
2318 * Record profiling data from argument profiling at an invoke with the
2319 * type system so that it can propagate it (speculation)
2320 *
2321 * @param dest_method target method for the call
2322 * @param bc what invoke bytecode is this?
2323 */
2324 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2325 if (!UseTypeSpeculation) {
2326 return;
2327 }
2328 const TypeFunc* tf = TypeFunc::make(dest_method);
2329 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2330 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2331 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2332 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2333 if (is_reference_type(targ->basic_type())) {
2334 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2335 ciKlass* better_type = nullptr;
2336 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2337 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2338 }
2339 i++;
2340 }
2341 }
2342 }
2343
2344 /**
2345 * Record profiling data from parameter profiling at an invoke with
2346 * the type system so that it can propagate it (speculation)
2347 */
2348 void GraphKit::record_profiled_parameters_for_speculation() {
2349 if (!UseTypeSpeculation) {
2350 return;
2351 }
2352 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2472 // The first null ends the list.
2473 Node* parm0, Node* parm1,
2474 Node* parm2, Node* parm3,
2475 Node* parm4, Node* parm5,
2476 Node* parm6, Node* parm7) {
2477 assert(call_addr != nullptr, "must not call null targets");
2478
2479 // Slow-path call
2480 bool is_leaf = !(flags & RC_NO_LEAF);
2481 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2482 if (call_name == nullptr) {
2483 assert(!is_leaf, "must supply name for leaf");
2484 call_name = OptoRuntime::stub_name(call_addr);
2485 }
2486 CallNode* call;
2487 if (!is_leaf) {
2488 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2489 } else if (flags & RC_NO_FP) {
2490 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2491 } else if (flags & RC_VECTOR){
2492 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2493 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2494 } else {
2495 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2496 }
2497
2498 // The following is similar to set_edges_for_java_call,
2499 // except that the memory effects of the call are restricted to AliasIdxRaw.
2500
2501 // Slow path call has no side-effects, uses few values
2502 bool wide_in = !(flags & RC_NARROW_MEM);
2503 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2504
2505 Node* prev_mem = nullptr;
2506 if (wide_in) {
2507 prev_mem = set_predefined_input_for_runtime_call(call);
2508 } else {
2509 assert(!wide_out, "narrow in => narrow out");
2510 Node* narrow_mem = memory(adr_type);
2511 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2512 }
2552
2553 if (has_io) {
2554 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2555 }
2556 return call;
2557
2558 }
2559
2560 // i2b
2561 Node* GraphKit::sign_extend_byte(Node* in) {
2562 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2563 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2564 }
2565
2566 // i2s
2567 Node* GraphKit::sign_extend_short(Node* in) {
2568 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2569 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2570 }
2571
2572 //------------------------------merge_memory-----------------------------------
2573 // Merge memory from one path into the current memory state.
2574 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2575 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2576 Node* old_slice = mms.force_memory();
2577 Node* new_slice = mms.memory2();
2578 if (old_slice != new_slice) {
2579 PhiNode* phi;
2580 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2581 if (mms.is_empty()) {
2582 // clone base memory Phi's inputs for this memory slice
2583 assert(old_slice == mms.base_memory(), "sanity");
2584 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2585 _gvn.set_type(phi, Type::MEMORY);
2586 for (uint i = 1; i < phi->req(); i++) {
2587 phi->init_req(i, old_slice->in(i));
2588 }
2589 } else {
2590 phi = old_slice->as_Phi(); // Phi was generated already
2591 }
2854
2855 // Now do a linear scan of the secondary super-klass array. Again, no real
2856 // performance impact (too rare) but it's gotta be done.
2857 // Since the code is rarely used, there is no penalty for moving it
2858 // out of line, and it can only improve I-cache density.
2859 // The decision to inline or out-of-line this final check is platform
2860 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2861 Node* psc = gvn.transform(
2862 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2863
2864 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2865 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2866 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2867
2868 // Return false path; set default control to true path.
2869 *ctrl = gvn.transform(r_ok_subtype);
2870 return gvn.transform(r_not_subtype);
2871 }
2872
2873 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2874 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2875 if (expand_subtype_check) {
2876 MergeMemNode* mem = merged_memory();
2877 Node* ctrl = control();
2878 Node* subklass = obj_or_subklass;
2879 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2880 subklass = load_object_klass(obj_or_subklass);
2881 }
2882
2883 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2884 set_control(ctrl);
2885 return n;
2886 }
2887
2888 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2889 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2890 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2891 set_control(_gvn.transform(new IfTrueNode(iff)));
2892 return _gvn.transform(new IfFalseNode(iff));
2893 }
2894
2895 // Profile-driven exact type check:
2896 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2897 float prob,
2898 Node* *casted_receiver) {
2899 assert(!klass->is_interface(), "no exact type check on interfaces");
2900
2901 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
2902 Node* recv_klass = load_object_klass(receiver);
2903 Node* want_klass = makecon(tklass);
2904 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2905 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2906 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2907 set_control( _gvn.transform(new IfTrueNode (iff)));
2908 Node* fail = _gvn.transform(new IfFalseNode(iff));
2909
2910 if (!stopped()) {
2911 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2912 const TypeOopPtr* recvx_type = tklass->as_instance_type();
2913 assert(recvx_type->klass_is_exact(), "");
2914
2915 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2916 // Subsume downstream occurrences of receiver with a cast to
2917 // recv_xtype, since now we know what the type will be.
2918 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2919 (*casted_receiver) = _gvn.transform(cast);
2920 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
2921 // (User must make the replace_in_map call.)
2922 }
2923 }
2924
2925 return fail;
2926 }
2927
2928 //------------------------------subtype_check_receiver-------------------------
2929 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2930 Node** casted_receiver) {
2931 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
2932 Node* want_klass = makecon(tklass);
2933
2934 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2935
2936 // Ignore interface type information until interface types are properly tracked.
2937 if (!stopped() && !klass->is_interface()) {
2938 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2939 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2940 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2941 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2942 (*casted_receiver) = _gvn.transform(cast);
2943 }
2944 }
2945
2946 return slow_ctl;
2947 }
2948
2949 //------------------------------seems_never_null-------------------------------
2950 // Use null_seen information if it is available from the profile.
2951 // If we see an unexpected null at a type check we record it and force a
2952 // recompile; the offending check will be recompiled to handle nulls.
2953 // If we see several offending BCIs, then all checks in the
2954 // method will be recompiled.
2955 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2956 speculating = !_gvn.type(obj)->speculative_maybe_null();
2957 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2958 if (UncommonNullCast // Cutout for this technique
2959 && obj != null() // And not the -Xcomp stupid case?
2960 && !too_many_traps(reason)
2961 ) {
2962 if (speculating) {
3031
3032 //------------------------maybe_cast_profiled_receiver-------------------------
3033 // If the profile has seen exactly one type, narrow to exactly that type.
3034 // Subsequent type checks will always fold up.
3035 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3036 const TypeKlassPtr* require_klass,
3037 ciKlass* spec_klass,
3038 bool safe_for_replace) {
3039 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3040
3041 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3042
3043 // Make sure we haven't already deoptimized from this tactic.
3044 if (too_many_traps_or_recompiles(reason))
3045 return nullptr;
3046
3047 // (No, this isn't a call, but it's enough like a virtual call
3048 // to use the same ciMethod accessor to get the profile info...)
3049 // If we have a speculative type use it instead of profiling (which
3050 // may not help us)
3051 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3052 if (exact_kls != nullptr) {// no cast failures here
3053 if (require_klass == nullptr ||
3054 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3055 // If we narrow the type to match what the type profile sees or
3056 // the speculative type, we can then remove the rest of the
3057 // cast.
3058 // This is a win, even if the exact_kls is very specific,
3059 // because downstream operations, such as method calls,
3060 // will often benefit from the sharper type.
3061 Node* exact_obj = not_null_obj; // will get updated in place...
3062 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3063 &exact_obj);
3064 { PreserveJVMState pjvms(this);
3065 set_control(slow_ctl);
3066 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3067 }
3068 if (safe_for_replace) {
3069 replace_in_map(not_null_obj, exact_obj);
3070 }
3071 return exact_obj;
3161 // If not_null_obj is dead, only null-path is taken
3162 if (stopped()) { // Doing instance-of on a null?
3163 set_control(null_ctl);
3164 return intcon(0);
3165 }
3166 region->init_req(_null_path, null_ctl);
3167 phi ->init_req(_null_path, intcon(0)); // Set null path value
3168 if (null_ctl == top()) {
3169 // Do this eagerly, so that pattern matches like is_diamond_phi
3170 // will work even during parsing.
3171 assert(_null_path == PATH_LIMIT-1, "delete last");
3172 region->del_req(_null_path);
3173 phi ->del_req(_null_path);
3174 }
3175
3176 // Do we know the type check always succeed?
3177 bool known_statically = false;
3178 if (_gvn.type(superklass)->singleton()) {
3179 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3180 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3181 if (subk->is_loaded()) {
3182 int static_res = C->static_subtype_check(superk, subk);
3183 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3184 }
3185 }
3186
3187 if (!known_statically) {
3188 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3189 // We may not have profiling here or it may not help us. If we
3190 // have a speculative type use it to perform an exact cast.
3191 ciKlass* spec_obj_type = obj_type->speculative_type();
3192 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3193 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3194 if (stopped()) { // Profile disagrees with this path.
3195 set_control(null_ctl); // Null is the only remaining possibility.
3196 return intcon(0);
3197 }
3198 if (cast_obj != nullptr) {
3199 not_null_obj = cast_obj;
3200 }
3201 }
3217 record_for_igvn(region);
3218
3219 // If we know the type check always succeeds then we don't use the
3220 // profiling data at this bytecode. Don't lose it, feed it to the
3221 // type system as a speculative type.
3222 if (safe_for_replace) {
3223 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3224 replace_in_map(obj, casted_obj);
3225 }
3226
3227 return _gvn.transform(phi);
3228 }
3229
3230 //-------------------------------gen_checkcast---------------------------------
3231 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3232 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3233 // uncommon-trap paths work. Adjust stack after this call.
3234 // If failure_control is supplied and not null, it is filled in with
3235 // the control edge for the cast failure. Otherwise, an appropriate
3236 // uncommon trap or exception is thrown.
3237 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3238 Node* *failure_control) {
3239 kill_dead_locals(); // Benefit all the uncommon traps
3240 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3241 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3242 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3243
3244 // Fast cutout: Check the case that the cast is vacuously true.
3245 // This detects the common cases where the test will short-circuit
3246 // away completely. We do this before we perform the null check,
3247 // because if the test is going to turn into zero code, we don't
3248 // want a residual null check left around. (Causes a slowdown,
3249 // for example, in some objArray manipulations, such as a[i]=a[j].)
3250 if (improved_klass_ptr_type->singleton()) {
3251 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3252 if (objtp != nullptr) {
3253 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3254 case Compile::SSC_always_true:
3255 // If we know the type check always succeed then we don't use
3256 // the profiling data at this bytecode. Don't lose it, feed it
3257 // to the type system as a speculative type.
3258 return record_profiled_receiver_for_speculation(obj);
3259 case Compile::SSC_always_false:
3260 // It needs a null check because a null will *pass* the cast check.
3261 // A non-null value will always produce an exception.
3262 if (!objtp->maybe_null()) {
3263 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3264 Deoptimization::DeoptReason reason = is_aastore ?
3265 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3266 builtin_throw(reason);
3267 return top();
3268 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3269 return null_assert(obj);
3270 }
3271 break; // Fall through to full check
3272 default:
3273 break;
3274 }
3275 }
3276 }
3277
3278 ciProfileData* data = nullptr;
3279 bool safe_for_replace = false;
3280 if (failure_control == nullptr) { // use MDO in regular case only
3281 assert(java_bc() == Bytecodes::_aastore ||
3282 java_bc() == Bytecodes::_checkcast,
3283 "interpreter profiles type checks only for these BCs");
3284 data = method()->method_data()->bci_to_data(bci());
3285 safe_for_replace = true;
3286 }
3287
3288 // Make the merge point
3289 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3290 RegionNode* region = new RegionNode(PATH_LIMIT);
3291 Node* phi = new PhiNode(region, toop);
3292 C->set_has_split_ifs(true); // Has chance for split-if optimization
3293
3294 // Use null-cast information if it is available
3295 bool speculative_not_null = false;
3296 bool never_see_null = ((failure_control == nullptr) // regular case only
3297 && seems_never_null(obj, data, speculative_not_null));
3298
3299 // Null check; get casted pointer; set region slot 3
3300 Node* null_ctl = top();
3301 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3302
3303 // If not_null_obj is dead, only null-path is taken
3304 if (stopped()) { // Doing instance-of on a null?
3305 set_control(null_ctl);
3306 return null();
3307 }
3308 region->init_req(_null_path, null_ctl);
3309 phi ->init_req(_null_path, null()); // Set null path value
3310 if (null_ctl == top()) {
3311 // Do this eagerly, so that pattern matches like is_diamond_phi
3312 // will work even during parsing.
3313 assert(_null_path == PATH_LIMIT-1, "delete last");
3314 region->del_req(_null_path);
3315 phi ->del_req(_null_path);
3316 }
3317
3318 Node* cast_obj = nullptr;
3319 if (improved_klass_ptr_type->klass_is_exact()) {
3320 // The following optimization tries to statically cast the speculative type of the object
3321 // (for example obtained during profiling) to the type of the superklass and then do a
3322 // dynamic check that the type of the object is what we expect. To work correctly
3323 // for checkcast and aastore the type of superklass should be exact.
3324 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3325 // We may not have profiling here or it may not help us. If we have
3326 // a speculative type use it to perform an exact cast.
3327 ciKlass* spec_obj_type = obj_type->speculative_type();
3328 if (spec_obj_type != nullptr || data != nullptr) {
3329 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3330 if (cast_obj != nullptr) {
3331 if (failure_control != nullptr) // failure is now impossible
3332 (*failure_control) = top();
3333 // adjust the type of the phi to the exact klass:
3334 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3335 }
3336 }
3337 }
3338
3339 if (cast_obj == nullptr) {
3340 // Generate the subtype check
3341 Node* improved_superklass = superklass;
3342 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3343 improved_superklass = makecon(improved_klass_ptr_type);
3344 }
3345 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3346
3347 // Plug in success path into the merge
3348 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3349 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3350 if (failure_control == nullptr) {
3351 if (not_subtype_ctrl != top()) { // If failure is possible
3352 PreserveJVMState pjvms(this);
3353 set_control(not_subtype_ctrl);
3354 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3355 Deoptimization::DeoptReason reason = is_aastore ?
3356 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3357 builtin_throw(reason);
3358 }
3359 } else {
3360 (*failure_control) = not_subtype_ctrl;
3361 }
3362 }
3363
3364 region->init_req(_obj_path, control());
3365 phi ->init_req(_obj_path, cast_obj);
3366
3367 // A merge of null or Casted-NotNull obj
3368 Node* res = _gvn.transform(phi);
3369
3370 // Note I do NOT always 'replace_in_map(obj,result)' here.
3371 // if( tk->klass()->can_be_primary_super() )
3372 // This means that if I successfully store an Object into an array-of-String
3373 // I 'forget' that the Object is really now known to be a String. I have to
3374 // do this because we don't have true union types for interfaces - if I store
3375 // a Baz into an array-of-Interface and then tell the optimizer it's an
3376 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3377 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3378 // replace_in_map( obj, res );
3379
3380 // Return final merged results
3381 set_control( _gvn.transform(region) );
3382 record_for_igvn(region);
3383
3384 return record_profiled_receiver_for_speculation(res);
3385 }
3386
3387 //------------------------------next_monitor-----------------------------------
3388 // What number should be given to the next monitor?
3389 int GraphKit::next_monitor() {
3390 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3391 int next = current + C->sync_stack_slots();
3392 // Keep the toplevel high water mark current:
3393 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3394 return current;
3395 }
3396
3397 //------------------------------insert_mem_bar---------------------------------
3398 // Memory barrier to avoid floating things around
3399 // The membar serves as a pinch point between both control and all memory slices.
3400 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3401 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3402 mb->init_req(TypeFunc::Control, control());
3403 mb->init_req(TypeFunc::Memory, reset_memory());
3404 Node* membar = _gvn.transform(mb);
3432 }
3433 Node* membar = _gvn.transform(mb);
3434 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3435 if (alias_idx == Compile::AliasIdxBot) {
3436 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3437 } else {
3438 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3439 }
3440 return membar;
3441 }
3442
3443 //------------------------------shared_lock------------------------------------
3444 // Emit locking code.
3445 FastLockNode* GraphKit::shared_lock(Node* obj) {
3446 // bci is either a monitorenter bc or InvocationEntryBci
3447 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3448 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3449
3450 if( !GenerateSynchronizationCode )
3451 return nullptr; // Not locking things?
3452 if (stopped()) // Dead monitor?
3453 return nullptr;
3454
3455 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3456
3457 // Box the stack location
3458 Node* box = new BoxLockNode(next_monitor());
3459 // Check for bailout after new BoxLockNode
3460 if (failing()) { return nullptr; }
3461 box = _gvn.transform(box);
3462 Node* mem = reset_memory();
3463
3464 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3465
3466 // Add monitor to debug info for the slow path. If we block inside the
3467 // slow path and de-opt, we need the monitor hanging around
3468 map()->push_monitor( flock );
3469
3470 const TypeFunc *tf = LockNode::lock_type();
3471 LockNode *lock = new LockNode(C, tf);
3500 }
3501 #endif
3502
3503 return flock;
3504 }
3505
3506
3507 //------------------------------shared_unlock----------------------------------
3508 // Emit unlocking code.
3509 void GraphKit::shared_unlock(Node* box, Node* obj) {
3510 // bci is either a monitorenter bc or InvocationEntryBci
3511 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3512 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3513
3514 if( !GenerateSynchronizationCode )
3515 return;
3516 if (stopped()) { // Dead monitor?
3517 map()->pop_monitor(); // Kill monitor from debug info
3518 return;
3519 }
3520
3521 // Memory barrier to avoid floating things down past the locked region
3522 insert_mem_bar(Op_MemBarReleaseLock);
3523
3524 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3525 UnlockNode *unlock = new UnlockNode(C, tf);
3526 #ifdef ASSERT
3527 unlock->set_dbg_jvms(sync_jvms());
3528 #endif
3529 uint raw_idx = Compile::AliasIdxRaw;
3530 unlock->init_req( TypeFunc::Control, control() );
3531 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3532 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3533 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3534 unlock->init_req( TypeFunc::ReturnAdr, top() );
3535
3536 unlock->init_req(TypeFunc::Parms + 0, obj);
3537 unlock->init_req(TypeFunc::Parms + 1, box);
3538 unlock = _gvn.transform(unlock)->as_Unlock();
3539
3540 Node* mem = reset_memory();
3541
3542 // unlock has no side-effects, sets few values
3543 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3544
3545 // Kill monitor from debug info
3546 map()->pop_monitor( );
3547 }
3548
3549 //-------------------------------get_layout_helper-----------------------------
3550 // If the given klass is a constant or known to be an array,
3551 // fetch the constant layout helper value into constant_value
3552 // and return null. Otherwise, load the non-constant
3553 // layout helper value, and return the node which represents it.
3554 // This two-faced routine is useful because allocation sites
3555 // almost always feature constant types.
3556 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3557 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3558 if (!StressReflectiveCode && klass_t != nullptr) {
3559 bool xklass = klass_t->klass_is_exact();
3560 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3561 jint lhelper;
3562 if (klass_t->isa_aryklassptr()) {
3563 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3564 if (is_reference_type(elem, true)) {
3565 elem = T_OBJECT;
3566 }
3567 lhelper = Klass::array_layout_helper(elem);
3568 } else {
3569 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3570 }
3571 if (lhelper != Klass::_lh_neutral_value) {
3572 constant_value = lhelper;
3573 return (Node*) nullptr;
3574 }
3575 }
3576 }
3577 constant_value = Klass::_lh_neutral_value; // put in a known value
3578 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3579 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3580 }
3581
3582 // We just put in an allocate/initialize with a big raw-memory effect.
3583 // Hook selected additional alias categories on the initialization.
3584 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3585 MergeMemNode* init_in_merge,
3586 Node* init_out_raw) {
3587 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3588 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3589
3590 Node* prevmem = kit.memory(alias_idx);
3591 init_in_merge->set_memory_at(alias_idx, prevmem);
3592 kit.set_memory(init_out_raw, alias_idx);
3593 }
3594
3595 //---------------------------set_output_for_allocation-------------------------
3596 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3597 const TypeOopPtr* oop_type,
3598 bool deoptimize_on_exception) {
3599 int rawidx = Compile::AliasIdxRaw;
3600 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3601 add_safepoint_edges(alloc);
3602 Node* allocx = _gvn.transform(alloc);
3603 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3604 // create memory projection for i_o
3605 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3606 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3607
3608 // create a memory projection as for the normal control path
3609 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3610 set_memory(malloc, rawidx);
3611
3612 // a normal slow-call doesn't change i_o, but an allocation does
3613 // we create a separate i_o projection for the normal control path
3614 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3615 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3616
3617 // put in an initialization barrier
3618 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3619 rawoop)->as_Initialize();
3620 assert(alloc->initialization() == init, "2-way macro link must work");
3621 assert(init ->allocation() == alloc, "2-way macro link must work");
3622 {
3623 // Extract memory strands which may participate in the new object's
3624 // initialization, and source them from the new InitializeNode.
3625 // This will allow us to observe initializations when they occur,
3626 // and link them properly (as a group) to the InitializeNode.
3627 assert(init->in(InitializeNode::Memory) == malloc, "");
3628 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3629 init->set_req(InitializeNode::Memory, minit_in);
3630 record_for_igvn(minit_in); // fold it up later, if possible
3631 Node* minit_out = memory(rawidx);
3632 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3633 // Add an edge in the MergeMem for the header fields so an access
3634 // to one of those has correct memory state
3635 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3636 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3637 if (oop_type->isa_aryptr()) {
3638 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3639 int elemidx = C->get_alias_index(telemref);
3640 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3641 } else if (oop_type->isa_instptr()) {
3642 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3643 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3644 ciField* field = ik->nonstatic_field_at(i);
3645 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3646 continue; // do not bother to track really large numbers of fields
3647 // Find (or create) the alias category for this field:
3648 int fieldidx = C->alias_type(field)->index();
3649 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3650 }
3651 }
3652 }
3653
3654 // Cast raw oop to the real thing...
3655 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3656 javaoop = _gvn.transform(javaoop);
3657 C->set_recent_alloc(control(), javaoop);
3658 assert(just_allocated_object(control()) == javaoop, "just allocated");
3659
3660 #ifdef ASSERT
3661 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3672 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3673 }
3674 }
3675 #endif //ASSERT
3676
3677 return javaoop;
3678 }
3679
3680 //---------------------------new_instance--------------------------------------
3681 // This routine takes a klass_node which may be constant (for a static type)
3682 // or may be non-constant (for reflective code). It will work equally well
3683 // for either, and the graph will fold nicely if the optimizer later reduces
3684 // the type to a constant.
3685 // The optional arguments are for specialized use by intrinsics:
3686 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3687 // - If 'return_size_val', report the total object size to the caller.
3688 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3689 Node* GraphKit::new_instance(Node* klass_node,
3690 Node* extra_slow_test,
3691 Node* *return_size_val,
3692 bool deoptimize_on_exception) {
3693 // Compute size in doublewords
3694 // The size is always an integral number of doublewords, represented
3695 // as a positive bytewise size stored in the klass's layout_helper.
3696 // The layout_helper also encodes (in a low bit) the need for a slow path.
3697 jint layout_con = Klass::_lh_neutral_value;
3698 Node* layout_val = get_layout_helper(klass_node, layout_con);
3699 int layout_is_con = (layout_val == nullptr);
3700
3701 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3702 // Generate the initial go-slow test. It's either ALWAYS (return a
3703 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3704 // case) a computed value derived from the layout_helper.
3705 Node* initial_slow_test = nullptr;
3706 if (layout_is_con) {
3707 assert(!StressReflectiveCode, "stress mode does not use these paths");
3708 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3709 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3710 } else { // reflective case
3711 // This reflective path is used by Unsafe.allocateInstance.
3712 // (It may be stress-tested by specifying StressReflectiveCode.)
3713 // Basically, we want to get into the VM is there's an illegal argument.
3714 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3715 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3716 if (extra_slow_test != intcon(0)) {
3717 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3718 }
3719 // (Macro-expander will further convert this to a Bool, if necessary.)
3730
3731 // Clear the low bits to extract layout_helper_size_in_bytes:
3732 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3733 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3734 size = _gvn.transform( new AndXNode(size, mask) );
3735 }
3736 if (return_size_val != nullptr) {
3737 (*return_size_val) = size;
3738 }
3739
3740 // This is a precise notnull oop of the klass.
3741 // (Actually, it need not be precise if this is a reflective allocation.)
3742 // It's what we cast the result to.
3743 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3744 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3745 const TypeOopPtr* oop_type = tklass->as_instance_type();
3746
3747 // Now generate allocation code
3748
3749 // The entire memory state is needed for slow path of the allocation
3750 // since GC and deoptimization can happened.
3751 Node *mem = reset_memory();
3752 set_all_memory(mem); // Create new memory state
3753
3754 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3755 control(), mem, i_o(),
3756 size, klass_node,
3757 initial_slow_test);
3758
3759 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3760 }
3761
3762 //-------------------------------new_array-------------------------------------
3763 // helper for both newarray and anewarray
3764 // The 'length' parameter is (obviously) the length of the array.
3765 // The optional arguments are for specialized use by intrinsics:
3766 // - If 'return_size_val', report the non-padded array size (sum of header size
3767 // and array body) to the caller.
3768 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3769 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3770 Node* length, // number of array elements
3771 int nargs, // number of arguments to push back for uncommon trap
3772 Node* *return_size_val,
3773 bool deoptimize_on_exception) {
3774 jint layout_con = Klass::_lh_neutral_value;
3775 Node* layout_val = get_layout_helper(klass_node, layout_con);
3776 int layout_is_con = (layout_val == nullptr);
3777
3778 if (!layout_is_con && !StressReflectiveCode &&
3779 !too_many_traps(Deoptimization::Reason_class_check)) {
3780 // This is a reflective array creation site.
3781 // Optimistically assume that it is a subtype of Object[],
3782 // so that we can fold up all the address arithmetic.
3783 layout_con = Klass::array_layout_helper(T_OBJECT);
3784 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3785 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3786 { BuildCutout unless(this, bol_lh, PROB_MAX);
3787 inc_sp(nargs);
3788 uncommon_trap(Deoptimization::Reason_class_check,
3789 Deoptimization::Action_maybe_recompile);
3790 }
3791 layout_val = nullptr;
3792 layout_is_con = true;
3793 }
3794
3795 // Generate the initial go-slow test. Make sure we do not overflow
3796 // if length is huge (near 2Gig) or negative! We do not need
3797 // exact double-words here, just a close approximation of needed
3798 // double-words. We can't add any offset or rounding bits, lest we
3799 // take a size -1 of bytes and make it positive. Use an unsigned
3800 // compare, so negative sizes look hugely positive.
3801 int fast_size_limit = FastAllocateSizeLimit;
3802 if (layout_is_con) {
3803 assert(!StressReflectiveCode, "stress mode does not use these paths");
3804 // Increase the size limit if we have exact knowledge of array type.
3805 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3806 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3807 }
3808
3809 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3810 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3811
3812 // --- Size Computation ---
3813 // array_size = round_to_heap(array_header + (length << elem_shift));
3814 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3815 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3816 // The rounding mask is strength-reduced, if possible.
3817 int round_mask = MinObjAlignmentInBytes - 1;
3818 Node* header_size = nullptr;
3819 // (T_BYTE has the weakest alignment and size restrictions...)
3820 if (layout_is_con) {
3821 int hsize = Klass::layout_helper_header_size(layout_con);
3822 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3823 if ((round_mask & ~right_n_bits(eshift)) == 0)
3824 round_mask = 0; // strength-reduce it if it goes away completely
3825 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3826 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3827 assert(header_size_min <= hsize, "generic minimum is smallest");
3828 header_size = intcon(hsize);
3829 } else {
3830 Node* hss = intcon(Klass::_lh_header_size_shift);
3831 Node* hsm = intcon(Klass::_lh_header_size_mask);
3832 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3833 header_size = _gvn.transform(new AndINode(header_size, hsm));
3834 }
3835
3836 Node* elem_shift = nullptr;
3837 if (layout_is_con) {
3838 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3839 if (eshift != 0)
3840 elem_shift = intcon(eshift);
3841 } else {
3842 // There is no need to mask or shift this value.
3843 // The semantics of LShiftINode include an implicit mask to 0x1F.
3844 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3845 elem_shift = layout_val;
3892 }
3893 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3894
3895 if (return_size_val != nullptr) {
3896 // This is the size
3897 (*return_size_val) = non_rounded_size;
3898 }
3899
3900 Node* size = non_rounded_size;
3901 if (round_mask != 0) {
3902 Node* mask1 = MakeConX(round_mask);
3903 size = _gvn.transform(new AddXNode(size, mask1));
3904 Node* mask2 = MakeConX(~round_mask);
3905 size = _gvn.transform(new AndXNode(size, mask2));
3906 }
3907 // else if round_mask == 0, the size computation is self-rounding
3908
3909 // Now generate allocation code
3910
3911 // The entire memory state is needed for slow path of the allocation
3912 // since GC and deoptimization can happened.
3913 Node *mem = reset_memory();
3914 set_all_memory(mem); // Create new memory state
3915
3916 if (initial_slow_test->is_Bool()) {
3917 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3918 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3919 }
3920
3921 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3922 Node* valid_length_test = _gvn.intcon(1);
3923 if (ary_type->isa_aryptr()) {
3924 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
3925 jint max = TypeAryPtr::max_array_length(bt);
3926 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
3927 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
3928 }
3929
3930 // Create the AllocateArrayNode and its result projections
3931 AllocateArrayNode* alloc
3932 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3933 control(), mem, i_o(),
3934 size, klass_node,
3935 initial_slow_test,
3936 length, valid_length_test);
3937
3938 // Cast to correct type. Note that the klass_node may be constant or not,
3939 // and in the latter case the actual array type will be inexact also.
3940 // (This happens via a non-constant argument to inline_native_newArray.)
3941 // In any case, the value of klass_node provides the desired array type.
3942 const TypeInt* length_type = _gvn.find_int_type(length);
3943 if (ary_type->isa_aryptr() && length_type != nullptr) {
3944 // Try to get a better type than POS for the size
3945 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3946 }
3947
3948 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3949
3950 array_ideal_length(alloc, ary_type, true);
3951 return javaoop;
3952 }
3953
3954 // The following "Ideal_foo" functions are placed here because they recognize
3955 // the graph shapes created by the functions immediately above.
3956
3957 //---------------------------Ideal_allocation----------------------------------
4065 set_all_memory(ideal.merged_memory());
4066 set_i_o(ideal.i_o());
4067 set_control(ideal.ctrl());
4068 }
4069
4070 void GraphKit::final_sync(IdealKit& ideal) {
4071 // Final sync IdealKit and graphKit.
4072 sync_kit(ideal);
4073 }
4074
4075 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4076 Node* len = load_array_length(load_String_value(str, set_ctrl));
4077 Node* coder = load_String_coder(str, set_ctrl);
4078 // Divide length by 2 if coder is UTF16
4079 return _gvn.transform(new RShiftINode(len, coder));
4080 }
4081
4082 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4083 int value_offset = java_lang_String::value_offset();
4084 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4085 false, nullptr, 0);
4086 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4087 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4088 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4089 ciTypeArrayKlass::make(T_BYTE), true, 0);
4090 Node* p = basic_plus_adr(str, str, value_offset);
4091 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4092 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4093 return load;
4094 }
4095
4096 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4097 if (!CompactStrings) {
4098 return intcon(java_lang_String::CODER_UTF16);
4099 }
4100 int coder_offset = java_lang_String::coder_offset();
4101 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4102 false, nullptr, 0);
4103 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4104
4105 Node* p = basic_plus_adr(str, str, coder_offset);
4106 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4107 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4108 return load;
4109 }
4110
4111 void GraphKit::store_String_value(Node* str, Node* value) {
4112 int value_offset = java_lang_String::value_offset();
4113 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4114 false, nullptr, 0);
4115 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4116
4117 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4118 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4119 }
4120
4121 void GraphKit::store_String_coder(Node* str, Node* value) {
4122 int coder_offset = java_lang_String::coder_offset();
4123 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4124 false, nullptr, 0);
4125 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4126
4127 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4128 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4129 }
4130
4131 // Capture src and dst memory state with a MergeMemNode
4132 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4133 if (src_type == dst_type) {
4134 // Types are equal, we don't need a MergeMemNode
4135 return memory(src_type);
4136 }
4137 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4138 record_for_igvn(merge); // fold it up later, if possible
4139 int src_idx = C->get_alias_index(src_type);
4140 int dst_idx = C->get_alias_index(dst_type);
4141 merge->set_memory_at(src_idx, memory(src_idx));
4142 merge->set_memory_at(dst_idx, memory(dst_idx));
4143 return merge;
4144 }
4217 i_char->init_req(2, AddI(i_char, intcon(2)));
4218
4219 set_control(IfFalse(iff));
4220 set_memory(st, TypeAryPtr::BYTES);
4221 }
4222
4223 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4224 if (!field->is_constant()) {
4225 return nullptr; // Field not marked as constant.
4226 }
4227 ciInstance* holder = nullptr;
4228 if (!field->is_static()) {
4229 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4230 if (const_oop != nullptr && const_oop->is_instance()) {
4231 holder = const_oop->as_instance();
4232 }
4233 }
4234 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4235 /*is_unsigned_load=*/false);
4236 if (con_type != nullptr) {
4237 return makecon(con_type);
4238 }
4239 return nullptr;
4240 }
4241
4242 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4243 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4244 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4245 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4246 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4247 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4248 return casted_obj;
4249 }
4250 return obj;
4251 }
|
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciFlatArrayKlass.hpp"
26 #include "ci/ciInlineKlass.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "asm/register.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/c2/barrierSetC2.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "oops/flatArrayKlass.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "utilities/bitMap.inline.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/growableArray.hpp"
57
58 //----------------------------GraphKit-----------------------------------------
59 // Main utility constructor.
60 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
61 : Phase(Phase::Parser),
62 _env(C->env()),
63 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
64 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
65 {
66 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
67 _exceptions = jvms->map()->next_exception();
68 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
69 set_jvms(jvms);
70 #ifdef ASSERT
71 if (_gvn.is_IterGVN() != nullptr) {
72 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
73 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
74 _worklist_size = _gvn.C->igvn_worklist()->size();
75 }
76 #endif
77 }
78
79 // Private constructor for parser.
80 GraphKit::GraphKit()
81 : Phase(Phase::Parser),
82 _env(C->env()),
83 _gvn(*C->initial_gvn()),
84 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
85 {
86 _exceptions = nullptr;
87 set_map(nullptr);
88 debug_only(_sp = -99);
89 debug_only(set_bci(-99));
90 }
91
92
93
94 //---------------------------clean_stack---------------------------------------
95 // Clear away rubbish from the stack area of the JVM state.
96 // This destroys any arguments that may be waiting on the stack.
341 }
342 static inline void add_one_req(Node* dstphi, Node* src) {
343 assert(is_hidden_merge(dstphi), "must be a special merge node");
344 assert(!is_hidden_merge(src), "must not be a special merge node");
345 dstphi->add_req(src);
346 }
347
348 //-----------------------combine_exception_states------------------------------
349 // This helper function combines exception states by building phis on a
350 // specially marked state-merging region. These regions and phis are
351 // untransformed, and can build up gradually. The region is marked by
352 // having a control input of its exception map, rather than null. Such
353 // regions do not appear except in this function, and in use_exception_state.
354 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
355 if (failing_internal()) {
356 return; // dying anyway...
357 }
358 JVMState* ex_jvms = ex_map->_jvms;
359 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
360 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
361 // TODO 8325632 Re-enable
362 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
363 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
364 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
365 assert(ex_map->req() == phi_map->req(), "matching maps");
366 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
367 Node* hidden_merge_mark = root();
368 Node* region = phi_map->control();
369 MergeMemNode* phi_mem = phi_map->merged_memory();
370 MergeMemNode* ex_mem = ex_map->merged_memory();
371 if (region->in(0) != hidden_merge_mark) {
372 // The control input is not (yet) a specially-marked region in phi_map.
373 // Make it so, and build some phis.
374 region = new RegionNode(2);
375 _gvn.set_type(region, Type::CONTROL);
376 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
377 region->init_req(1, phi_map->control());
378 phi_map->set_control(region);
379 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
380 record_for_igvn(io_phi);
381 _gvn.set_type(io_phi, Type::ABIO);
382 phi_map->set_i_o(io_phi);
870 if (PrintMiscellaneous && (Verbose || WizardMode)) {
871 tty->print_cr("Zombie local %d: ", local);
872 jvms->dump();
873 }
874 return false;
875 }
876 }
877 }
878 return true;
879 }
880
881 #endif //ASSERT
882
883 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
884 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
885 ciMethod* cur_method = jvms->method();
886 int cur_bci = jvms->bci();
887 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
888 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
889 return Interpreter::bytecode_should_reexecute(code) ||
890 (is_anewarray && (code == Bytecodes::_multianewarray));
891 // Reexecute _multianewarray bytecode which was replaced with
892 // sequence of [a]newarray. See Parse::do_multianewarray().
893 //
894 // Note: interpreter should not have it set since this optimization
895 // is limited by dimensions and guarded by flag so in some cases
896 // multianewarray() runtime calls will be generated and
897 // the bytecode should not be reexecutes (stack will not be reset).
898 } else {
899 return false;
900 }
901 }
902
903 // Helper function for adding JVMState and debug information to node
904 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
905 // Add the safepoint edges to the call (or other safepoint).
906
907 // Make sure dead locals are set to top. This
908 // should help register allocation time and cut down on the size
909 // of the deoptimization information.
910 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
961 }
962
963 // Presize the call:
964 DEBUG_ONLY(uint non_debug_edges = call->req());
965 call->add_req_batch(top(), youngest_jvms->debug_depth());
966 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
967
968 // Set up edges so that the call looks like this:
969 // Call [state:] ctl io mem fptr retadr
970 // [parms:] parm0 ... parmN
971 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
972 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
973 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
974 // Note that caller debug info precedes callee debug info.
975
976 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
977 uint debug_ptr = call->req();
978
979 // Loop over the map input edges associated with jvms, add them
980 // to the call node, & reset all offsets to match call node array.
981
982 JVMState* callee_jvms = nullptr;
983 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
984 uint debug_end = debug_ptr;
985 uint debug_start = debug_ptr - in_jvms->debug_size();
986 debug_ptr = debug_start; // back up the ptr
987
988 uint p = debug_start; // walks forward in [debug_start, debug_end)
989 uint j, k, l;
990 SafePointNode* in_map = in_jvms->map();
991 out_jvms->set_map(call);
992
993 if (can_prune_locals) {
994 assert(in_jvms->method() == out_jvms->method(), "sanity");
995 // If the current throw can reach an exception handler in this JVMS,
996 // then we must keep everything live that can reach that handler.
997 // As a quick and dirty approximation, we look for any handlers at all.
998 if (in_jvms->method()->has_exception_handlers()) {
999 can_prune_locals = false;
1000 }
1001 }
1002
1003 // Add the Locals
1004 k = in_jvms->locoff();
1005 l = in_jvms->loc_size();
1006 out_jvms->set_locoff(p);
1007 if (!can_prune_locals) {
1008 for (j = 0; j < l; j++) {
1009 call->set_req(p++, in_map->in(k + j));
1010 }
1011 } else {
1012 p += l; // already set to top above by add_req_batch
1013 }
1014
1015 // Add the Expression Stack
1016 k = in_jvms->stkoff();
1017 l = in_jvms->sp();
1018 out_jvms->set_stkoff(p);
1019 if (!can_prune_locals) {
1020 for (j = 0; j < l; j++) {
1021 call->set_req(p++, in_map->in(k + j));
1022 }
1023 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1024 // Divide stack into {S0,...,S1}, where S0 is set to top.
1025 uint s1 = stack_slots_not_pruned;
1026 stack_slots_not_pruned = 0; // for next iteration
1027 if (s1 > l) s1 = l;
1028 uint s0 = l - s1;
1029 p += s0; // skip the tops preinstalled by add_req_batch
1030 for (j = s0; j < l; j++)
1031 call->set_req(p++, in_map->in(k+j));
1032 } else {
1033 p += l; // already set to top above by add_req_batch
1034 }
1035
1036 // Add the Monitors
1037 k = in_jvms->monoff();
1038 l = in_jvms->mon_size();
1039 out_jvms->set_monoff(p);
1040 for (j = 0; j < l; j++)
1041 call->set_req(p++, in_map->in(k+j));
1042
1043 // Copy any scalar object fields.
1044 k = in_jvms->scloff();
1045 l = in_jvms->scl_size();
1046 out_jvms->set_scloff(p);
1047 for (j = 0; j < l; j++)
1048 call->set_req(p++, in_map->in(k+j));
1049
1050 // Finish the new jvms.
1051 out_jvms->set_endoff(p);
1052
1053 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1054 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1055 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1056 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1057 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1058 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1059
1060 // Update the two tail pointers in parallel.
1061 callee_jvms = out_jvms;
1062 out_jvms = out_jvms->caller();
1063 in_jvms = in_jvms->caller();
1064 }
1065
1066 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1067
1068 // Test the correctness of JVMState::debug_xxx accessors:
1069 assert(call->jvms()->debug_start() == non_debug_edges, "");
1070 assert(call->jvms()->debug_end() == call->req(), "");
1071 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1072 }
1073
1074 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1075 Bytecodes::Code code = java_bc();
1076 if (code == Bytecodes::_wide) {
1077 code = method()->java_code_at_bci(bci() + 1);
1078 }
1079
1080 if (code != Bytecodes::_illegal) {
1081 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1212 Node* GraphKit::ConvI2UL(Node* offset) {
1213 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1214 if (offset_con != (juint) Type::OffsetBot) {
1215 return longcon((julong) offset_con);
1216 }
1217 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1218 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1219 return _gvn.transform( new AndLNode(conv, mask) );
1220 }
1221
1222 Node* GraphKit::ConvL2I(Node* offset) {
1223 // short-circuit a common case
1224 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1225 if (offset_con != (jlong)Type::OffsetBot) {
1226 return intcon((int) offset_con);
1227 }
1228 return _gvn.transform( new ConvL2INode(offset));
1229 }
1230
1231 //-------------------------load_object_klass-----------------------------------
1232 Node* GraphKit::load_object_klass(Node* obj, bool fold_for_arrays) {
1233 // Special-case a fresh allocation to avoid building nodes:
1234 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1235 if (akls != nullptr) return akls;
1236 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1237 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT, fold_for_arrays));
1238 }
1239
1240 //-------------------------load_array_length-----------------------------------
1241 Node* GraphKit::load_array_length(Node* array) {
1242 // Special-case a fresh allocation to avoid building nodes:
1243 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1244 Node *alen;
1245 if (alloc == nullptr) {
1246 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1247 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1248 } else {
1249 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1250 }
1251 return alen;
1252 }
1253
1254 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1255 const TypeOopPtr* oop_type,
1256 bool replace_length_in_map) {
1257 Node* length = alloc->Ideal_length();
1266 replace_in_map(length, ccast);
1267 }
1268 return ccast;
1269 }
1270 }
1271 return length;
1272 }
1273
1274 //------------------------------do_null_check----------------------------------
1275 // Helper function to do a null pointer check. Returned value is
1276 // the incoming address with null casted away. You are allowed to use the
1277 // not-null value only if you are control dependent on the test.
1278 #ifndef PRODUCT
1279 extern uint explicit_null_checks_inserted,
1280 explicit_null_checks_elided;
1281 #endif
1282 Node* GraphKit::null_check_common(Node* value, BasicType type,
1283 // optional arguments for variations:
1284 bool assert_null,
1285 Node* *null_control,
1286 bool speculative,
1287 bool is_init_check) {
1288 assert(!assert_null || null_control == nullptr, "not both at once");
1289 if (stopped()) return top();
1290 NOT_PRODUCT(explicit_null_checks_inserted++);
1291
1292 if (value->is_InlineType()) {
1293 // Null checking a scalarized but nullable inline type. Check the IsInit
1294 // input instead of the oop input to avoid keeping buffer allocations alive.
1295 InlineTypeNode* vtptr = value->as_InlineType();
1296 while (vtptr->get_oop()->is_InlineType()) {
1297 vtptr = vtptr->get_oop()->as_InlineType();
1298 }
1299 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1300 if (stopped()) {
1301 return top();
1302 }
1303 if (assert_null) {
1304 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1305 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1306 // replace_in_map(value, vtptr);
1307 // return vtptr;
1308 replace_in_map(value, null());
1309 return null();
1310 }
1311 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1312 return cast_not_null(value, do_replace_in_map);
1313 }
1314
1315 // Construct null check
1316 Node *chk = nullptr;
1317 switch(type) {
1318 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1319 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1320 case T_ARRAY : // fall through
1321 type = T_OBJECT; // simplify further tests
1322 case T_OBJECT : {
1323 const Type *t = _gvn.type( value );
1324
1325 const TypeOopPtr* tp = t->isa_oopptr();
1326 if (tp != nullptr && !tp->is_loaded()
1327 // Only for do_null_check, not any of its siblings:
1328 && !assert_null && null_control == nullptr) {
1329 // Usually, any field access or invocation on an unloaded oop type
1330 // will simply fail to link, since the statically linked class is
1331 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1332 // the static class is loaded but the sharper oop type is not.
1333 // Rather than checking for this obscure case in lots of places,
1334 // we simply observe that a null check on an unloaded class
1398 }
1399 Node *oldcontrol = control();
1400 set_control(cfg);
1401 Node *res = cast_not_null(value);
1402 set_control(oldcontrol);
1403 NOT_PRODUCT(explicit_null_checks_elided++);
1404 return res;
1405 }
1406 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1407 if (cfg == nullptr) break; // Quit at region nodes
1408 depth++;
1409 }
1410 }
1411
1412 //-----------
1413 // Branch to failure if null
1414 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1415 Deoptimization::DeoptReason reason;
1416 if (assert_null) {
1417 reason = Deoptimization::reason_null_assert(speculative);
1418 } else if (type == T_OBJECT || is_init_check) {
1419 reason = Deoptimization::reason_null_check(speculative);
1420 } else {
1421 reason = Deoptimization::Reason_div0_check;
1422 }
1423 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1424 // ciMethodData::has_trap_at will return a conservative -1 if any
1425 // must-be-null assertion has failed. This could cause performance
1426 // problems for a method after its first do_null_assert failure.
1427 // Consider using 'Reason_class_check' instead?
1428
1429 // To cause an implicit null check, we set the not-null probability
1430 // to the maximum (PROB_MAX). For an explicit check the probability
1431 // is set to a smaller value.
1432 if (null_control != nullptr || too_many_traps(reason)) {
1433 // probability is less likely
1434 ok_prob = PROB_LIKELY_MAG(3);
1435 } else if (!assert_null &&
1436 (ImplicitNullCheckThreshold > 0) &&
1437 method() != nullptr &&
1438 (method()->method_data()->trap_count(reason)
1472 }
1473
1474 if (assert_null) {
1475 // Cast obj to null on this path.
1476 replace_in_map(value, zerocon(type));
1477 return zerocon(type);
1478 }
1479
1480 // Cast obj to not-null on this path, if there is no null_control.
1481 // (If there is a null_control, a non-null value may come back to haunt us.)
1482 if (type == T_OBJECT) {
1483 Node* cast = cast_not_null(value, false);
1484 if (null_control == nullptr || (*null_control) == top())
1485 replace_in_map(value, cast);
1486 value = cast;
1487 }
1488
1489 return value;
1490 }
1491
1492 //------------------------------cast_not_null----------------------------------
1493 // Cast obj to not-null on this path
1494 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1495 if (obj->is_InlineType()) {
1496 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1497 vt->as_InlineType()->set_is_init(_gvn);
1498 vt = _gvn.transform(vt);
1499 if (do_replace_in_map) {
1500 replace_in_map(obj, vt);
1501 }
1502 return vt;
1503 }
1504 const Type *t = _gvn.type(obj);
1505 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1506 // Object is already not-null?
1507 if( t == t_not_null ) return obj;
1508
1509 Node* cast = new CastPPNode(control(), obj,t_not_null);
1510 cast = _gvn.transform( cast );
1511
1512 // Scan for instances of 'obj' in the current JVM mapping.
1513 // These instances are known to be not-null after the test.
1514 if (do_replace_in_map)
1515 replace_in_map(obj, cast);
1516
1517 return cast; // Return casted value
1518 }
1519
1520 Node* GraphKit::cast_to_non_larval(Node* obj) {
1521 const Type* obj_type = gvn().type(obj);
1522 if (obj->is_InlineType() || !obj_type->is_inlinetypeptr()) {
1523 return obj;
1524 }
1525
1526 Node* new_obj = InlineTypeNode::make_from_oop(this, obj, obj_type->inline_klass());
1527 replace_in_map(obj, new_obj);
1528 return new_obj;
1529 }
1530
1531 // Sometimes in intrinsics, we implicitly know an object is not null
1532 // (there's no actual null check) so we can cast it to not null. In
1533 // the course of optimizations, the input to the cast can become null.
1534 // In that case that data path will die and we need the control path
1535 // to become dead as well to keep the graph consistent. So we have to
1536 // add a check for null for which one branch can't be taken. It uses
1537 // an OpaqueNotNull node that will cause the check to be removed after loop
1538 // opts so the test goes away and the compiled code doesn't execute a
1539 // useless check.
1540 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1541 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1542 return value;
1543 }
1544 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1545 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1546 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1547 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1548 _gvn.set_type(iff, iff->Value(&_gvn));
1549 if (!tst->is_Con()) {
1550 record_for_igvn(iff);
1623 // These are layered on top of the factory methods in LoadNode and StoreNode,
1624 // and integrate with the parser's memory state and _gvn engine.
1625 //
1626
1627 // factory methods in "int adr_idx"
1628 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1629 MemNode::MemOrd mo,
1630 LoadNode::ControlDependency control_dependency,
1631 bool require_atomic_access,
1632 bool unaligned,
1633 bool mismatched,
1634 bool unsafe,
1635 uint8_t barrier_data) {
1636 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1637 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1638 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1639 debug_only(adr_type = C->get_adr_type(adr_idx));
1640 Node* mem = memory(adr_idx);
1641 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1642 ld = _gvn.transform(ld);
1643
1644 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1645 // Improve graph before escape analysis and boxing elimination.
1646 record_for_igvn(ld);
1647 if (ld->is_DecodeN()) {
1648 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1649 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1650 // a Phi). Recording such cases is still perfectly sound, but may be
1651 // unnecessary and result in some minor IGVN overhead.
1652 record_for_igvn(ld->in(1));
1653 }
1654 }
1655 return ld;
1656 }
1657
1658 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1659 MemNode::MemOrd mo,
1660 bool require_atomic_access,
1661 bool unaligned,
1662 bool mismatched,
1663 bool unsafe,
1677 if (unsafe) {
1678 st->as_Store()->set_unsafe_access();
1679 }
1680 st->as_Store()->set_barrier_data(barrier_data);
1681 st = _gvn.transform(st);
1682 set_memory(st, adr_idx);
1683 // Back-to-back stores can only remove intermediate store with DU info
1684 // so push on worklist for optimizer.
1685 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1686 record_for_igvn(st);
1687
1688 return st;
1689 }
1690
1691 Node* GraphKit::access_store_at(Node* obj,
1692 Node* adr,
1693 const TypePtr* adr_type,
1694 Node* val,
1695 const Type* val_type,
1696 BasicType bt,
1697 DecoratorSet decorators,
1698 bool safe_for_replace,
1699 const InlineTypeNode* vt) {
1700 // Transformation of a value which could be null pointer (CastPP #null)
1701 // could be delayed during Parse (for example, in adjust_map_after_if()).
1702 // Execute transformation here to avoid barrier generation in such case.
1703 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1704 val = _gvn.makecon(TypePtr::NULL_PTR);
1705 }
1706
1707 if (stopped()) {
1708 return top(); // Dead path ?
1709 }
1710
1711 assert(val != nullptr, "not dead path");
1712 if (val->is_InlineType()) {
1713 // Store to non-flat field. Buffer the inline type and make sure
1714 // the store is re-executed if the allocation triggers deoptimization.
1715 PreserveReexecuteState preexecs(this);
1716 jvms()->set_should_reexecute(true);
1717 val = val->as_InlineType()->buffer(this, safe_for_replace);
1718 }
1719
1720 C2AccessValuePtr addr(adr, adr_type);
1721 C2AccessValue value(val, val_type);
1722 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1723 if (access.is_raw()) {
1724 return _barrier_set->BarrierSetC2::store_at(access, value);
1725 } else {
1726 return _barrier_set->store_at(access, value);
1727 }
1728 }
1729
1730 Node* GraphKit::access_load_at(Node* obj, // containing obj
1731 Node* adr, // actual address to store val at
1732 const TypePtr* adr_type,
1733 const Type* val_type,
1734 BasicType bt,
1735 DecoratorSet decorators,
1736 Node* ctl) {
1737 if (stopped()) {
1738 return top(); // Dead path ?
1739 }
1740
1741 C2AccessValuePtr addr(adr, adr_type);
1742 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1743 if (access.is_raw()) {
1744 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1745 } else {
1746 return _barrier_set->load_at(access, val_type);
1747 }
1748 }
1749
1750 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1751 const Type* val_type,
1752 BasicType bt,
1753 DecoratorSet decorators) {
1754 if (stopped()) {
1755 return top(); // Dead path ?
1756 }
1757
1758 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1759 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1760 if (access.is_raw()) {
1761 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1762 } else {
1827 Node* new_val,
1828 const Type* value_type,
1829 BasicType bt,
1830 DecoratorSet decorators) {
1831 C2AccessValuePtr addr(adr, adr_type);
1832 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1833 if (access.is_raw()) {
1834 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1835 } else {
1836 return _barrier_set->atomic_add_at(access, new_val, value_type);
1837 }
1838 }
1839
1840 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1841 return _barrier_set->clone(this, src, dst, size, is_array);
1842 }
1843
1844 //-------------------------array_element_address-------------------------
1845 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1846 const TypeInt* sizetype, Node* ctrl) {
1847 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1848 uint shift;
1849 if (arytype->is_flat() && arytype->klass_is_exact()) {
1850 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1851 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1852 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1853 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1854 // though we don't need the address node in this case and throw it away again.
1855 shift = arytype->flat_log_elem_size();
1856 } else {
1857 shift = exact_log2(type2aelembytes(elembt));
1858 }
1859 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1860
1861 // short-circuit a common case (saves lots of confusing waste motion)
1862 jint idx_con = find_int_con(idx, -1);
1863 if (idx_con >= 0) {
1864 intptr_t offset = header + ((intptr_t)idx_con << shift);
1865 return basic_plus_adr(ary, offset);
1866 }
1867
1868 // must be correct type for alignment purposes
1869 Node* base = basic_plus_adr(ary, header);
1870 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1871 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1872 return basic_plus_adr(ary, base, scale);
1873 }
1874
1875 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* vk, bool is_null_free, bool is_not_null_free, bool is_atomic) {
1876 assert(vk->maybe_flat_in_array(), "element of type %s cannot be flat in array", vk->name()->as_utf8());
1877 if (!vk->has_nullable_atomic_layout()) {
1878 // Element does not have a nullable flat layout, cannot be nullable
1879 is_null_free = true;
1880 }
1881 if (!vk->has_atomic_layout() && !vk->has_non_atomic_layout()) {
1882 // Element does not have a null-free flat layout, cannot be null-free
1883 is_not_null_free = true;
1884 }
1885 if (is_null_free) {
1886 // TODO 8350865 Impossible type
1887 is_not_null_free = false;
1888 }
1889
1890 bool is_exact = is_null_free || is_not_null_free;
1891 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* flat */ true, is_null_free, is_atomic);
1892 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1893 arytype = arytype->cast_to_exactness(is_exact);
1894 arytype = arytype->cast_to_not_null_free(is_not_null_free);
1895 return _gvn.transform(new CastPPNode(control(), array, arytype, ConstraintCastNode::StrongDependency));
1896 }
1897
1898 //-------------------------load_array_element-------------------------
1899 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1900 const Type* elemtype = arytype->elem();
1901 BasicType elembt = elemtype->array_element_basic_type();
1902 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1903 if (elembt == T_NARROWOOP) {
1904 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1905 }
1906 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1907 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1908 return ld;
1909 }
1910
1911 //-------------------------set_arguments_for_java_call-------------------------
1912 // Arguments (pre-popped from the stack) are taken from the JVMS.
1913 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1914 PreserveReexecuteState preexecs(this);
1915 if (EnableValhalla) {
1916 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1917 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1918 jvms()->set_should_reexecute(true);
1919 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1920 inc_sp(arg_size);
1921 }
1922 // Add the call arguments
1923 const TypeTuple* domain = call->tf()->domain_sig();
1924 uint nargs = domain->cnt();
1925 int arg_num = 0;
1926 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1927 Node* arg = argument(i-TypeFunc::Parms);
1928 const Type* t = domain->field_at(i);
1929 // TODO 8284443 A static call to a mismatched method should still be scalarized
1930 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1931 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1932 if (!arg->is_InlineType()) {
1933 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1934 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1935 }
1936 InlineTypeNode* vt = arg->as_InlineType();
1937 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1938 // If an inline type argument is passed as fields, attach the Method* to the call site
1939 // to be able to access the extended signature later via attached_method_before_pc().
1940 // For example, see CompiledMethod::preserve_callee_argument_oops().
1941 call->set_override_symbolic_info(true);
1942 // Register an evol dependency on the callee method to make sure that this method is deoptimized and
1943 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1944 C->dependencies()->assert_evol_method(call->method());
1945 arg_num++;
1946 continue;
1947 } else if (arg->is_InlineType()) {
1948 // Pass inline type argument via oop to callee
1949 arg = arg->as_InlineType()->buffer(this, true);
1950 }
1951 if (t != Type::HALF) {
1952 arg_num++;
1953 }
1954 call->init_req(idx++, arg);
1955 }
1956 }
1957
1958 //---------------------------set_edges_for_java_call---------------------------
1959 // Connect a newly created call into the current JVMS.
1960 // A return value node (if any) is returned from set_edges_for_java_call.
1961 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1962
1963 // Add the predefined inputs:
1964 call->init_req( TypeFunc::Control, control() );
1965 call->init_req( TypeFunc::I_O , i_o() );
1966 call->init_req( TypeFunc::Memory , reset_memory() );
1967 call->init_req( TypeFunc::FramePtr, frameptr() );
1968 call->init_req( TypeFunc::ReturnAdr, top() );
1969
1970 add_safepoint_edges(call, must_throw);
1971
1972 Node* xcall = _gvn.transform(call);
1973
1974 if (xcall == top()) {
1975 set_control(top());
1976 return;
1977 }
1978 assert(xcall == call, "call identity is stable");
1979
1980 // Re-use the current map to produce the result.
1981
1982 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1983 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1984 set_all_memory_call(xcall, separate_io_proj);
1985
1986 //return xcall; // no need, caller already has it
1987 }
1988
1989 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1990 if (stopped()) return top(); // maybe the call folded up?
1991
1992 // Note: Since any out-of-line call can produce an exception,
1993 // we always insert an I_O projection from the call into the result.
1994
1995 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1996
1997 if (separate_io_proj) {
1998 // The caller requested separate projections be used by the fall
1999 // through and exceptional paths, so replace the projections for
2000 // the fall through path.
2001 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2002 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2003 }
2004
2005 // Capture the return value, if any.
2006 Node* ret;
2007 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2008 ret = top();
2009 } else if (call->tf()->returns_inline_type_as_fields()) {
2010 // Return of multiple values (inline type fields): we create a
2011 // InlineType node, each field is a projection from the call.
2012 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2013 uint base_input = TypeFunc::Parms;
2014 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2015 } else {
2016 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2017 ciType* t = call->method()->return_type();
2018 if (t->is_klass()) {
2019 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2020 if (type->is_inlinetypeptr()) {
2021 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2022 }
2023 }
2024 }
2025
2026 return ret;
2027 }
2028
2029 //--------------------set_predefined_input_for_runtime_call--------------------
2030 // Reading and setting the memory state is way conservative here.
2031 // The real problem is that I am not doing real Type analysis on memory,
2032 // so I cannot distinguish card mark stores from other stores. Across a GC
2033 // point the Store Barrier and the card mark memory has to agree. I cannot
2034 // have a card mark store and its barrier split across the GC point from
2035 // either above or below. Here I get that to happen by reading ALL of memory.
2036 // A better answer would be to separate out card marks from other memory.
2037 // For now, return the input memory state, so that it can be reused
2038 // after the call, if this call has restricted memory effects.
2039 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2040 // Set fixed predefined input arguments
2041 Node* memory = reset_memory();
2042 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2043 call->init_req( TypeFunc::Control, control() );
2044 call->init_req( TypeFunc::I_O, top() ); // does no i/o
2045 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
2096 if (use->is_MergeMem()) {
2097 wl.push(use);
2098 }
2099 }
2100 }
2101
2102 // Replace the call with the current state of the kit.
2103 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2104 JVMState* ejvms = nullptr;
2105 if (has_exceptions()) {
2106 ejvms = transfer_exceptions_into_jvms();
2107 }
2108
2109 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2110 ReplacedNodes replaced_nodes_exception;
2111 Node* ex_ctl = top();
2112
2113 SafePointNode* final_state = stop();
2114
2115 // Find all the needed outputs of this call
2116 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2117
2118 Unique_Node_List wl;
2119 Node* init_mem = call->in(TypeFunc::Memory);
2120 Node* final_mem = final_state->in(TypeFunc::Memory);
2121 Node* final_ctl = final_state->in(TypeFunc::Control);
2122 Node* final_io = final_state->in(TypeFunc::I_O);
2123
2124 // Replace all the old call edges with the edges from the inlining result
2125 if (callprojs->fallthrough_catchproj != nullptr) {
2126 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2127 }
2128 if (callprojs->fallthrough_memproj != nullptr) {
2129 if (final_mem->is_MergeMem()) {
2130 // Parser's exits MergeMem was not transformed but may be optimized
2131 final_mem = _gvn.transform(final_mem);
2132 }
2133 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2134 add_mergemem_users_to_worklist(wl, final_mem);
2135 }
2136 if (callprojs->fallthrough_ioproj != nullptr) {
2137 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2138 }
2139
2140 // Replace the result with the new result if it exists and is used
2141 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2142 // If the inlined code is dead, the result projections for an inline type returned as
2143 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2144 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2145 "unexpected number of results");
2146 C->gvn_replace_by(callprojs->resproj[0], result);
2147 }
2148
2149 if (ejvms == nullptr) {
2150 // No exception edges to simply kill off those paths
2151 if (callprojs->catchall_catchproj != nullptr) {
2152 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2153 }
2154 if (callprojs->catchall_memproj != nullptr) {
2155 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2156 }
2157 if (callprojs->catchall_ioproj != nullptr) {
2158 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2159 }
2160 // Replace the old exception object with top
2161 if (callprojs->exobj != nullptr) {
2162 C->gvn_replace_by(callprojs->exobj, C->top());
2163 }
2164 } else {
2165 GraphKit ekit(ejvms);
2166
2167 // Load my combined exception state into the kit, with all phis transformed:
2168 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2169 replaced_nodes_exception = ex_map->replaced_nodes();
2170
2171 Node* ex_oop = ekit.use_exception_state(ex_map);
2172
2173 if (callprojs->catchall_catchproj != nullptr) {
2174 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2175 ex_ctl = ekit.control();
2176 }
2177 if (callprojs->catchall_memproj != nullptr) {
2178 Node* ex_mem = ekit.reset_memory();
2179 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2180 add_mergemem_users_to_worklist(wl, ex_mem);
2181 }
2182 if (callprojs->catchall_ioproj != nullptr) {
2183 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2184 }
2185
2186 // Replace the old exception object with the newly created one
2187 if (callprojs->exobj != nullptr) {
2188 C->gvn_replace_by(callprojs->exobj, ex_oop);
2189 }
2190 }
2191
2192 // Disconnect the call from the graph
2193 call->disconnect_inputs(C);
2194 C->gvn_replace_by(call, C->top());
2195
2196 // Clean up any MergeMems that feed other MergeMems since the
2197 // optimizer doesn't like that.
2198 while (wl.size() > 0) {
2199 _gvn.transform(wl.pop());
2200 }
2201
2202 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2203 replaced_nodes.apply(C, final_ctl);
2204 }
2205 if (!ex_ctl->is_top() && do_replaced_nodes) {
2206 replaced_nodes_exception.apply(C, ex_ctl);
2207 }
2208 }
2209
2210
2211 //------------------------------increment_counter------------------------------
2212 // for statistics: increment a VM counter by 1
2213
2214 void GraphKit::increment_counter(address counter_addr) {
2215 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2216 increment_counter(adr1);
2217 }
2218
2219 void GraphKit::increment_counter(Node* counter_addr) {
2220 Node* ctrl = control();
2221 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2222 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2382 *
2383 * @param n node that the type applies to
2384 * @param exact_kls type from profiling
2385 * @param maybe_null did profiling see null?
2386 *
2387 * @return node with improved type
2388 */
2389 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2390 const Type* current_type = _gvn.type(n);
2391 assert(UseTypeSpeculation, "type speculation must be on");
2392
2393 const TypePtr* speculative = current_type->speculative();
2394
2395 // Should the klass from the profile be recorded in the speculative type?
2396 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2397 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2398 const TypeOopPtr* xtype = tklass->as_instance_type();
2399 assert(xtype->klass_is_exact(), "Should be exact");
2400 // Any reason to believe n is not null (from this profiling or a previous one)?
2401 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2402 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2403 // record the new speculative type's depth
2404 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2405 speculative = speculative->with_inline_depth(jvms()->depth());
2406 } else if (current_type->would_improve_ptr(ptr_kind)) {
2407 // Profiling report that null was never seen so we can change the
2408 // speculative type to non null ptr.
2409 if (ptr_kind == ProfileAlwaysNull) {
2410 speculative = TypePtr::NULL_PTR;
2411 } else {
2412 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2413 const TypePtr* ptr = TypePtr::NOTNULL;
2414 if (speculative != nullptr) {
2415 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2416 } else {
2417 speculative = ptr;
2418 }
2419 }
2420 }
2421
2422 if (speculative != current_type->speculative()) {
2423 // Build a type with a speculative type (what we think we know
2424 // about the type but will need a guard when we use it)
2425 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2426 // We're changing the type, we need a new CheckCast node to carry
2427 // the new type. The new type depends on the control: what
2428 // profiling tells us is only valid from here as far as we can
2429 // tell.
2430 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2431 cast = _gvn.transform(cast);
2432 replace_in_map(n, cast);
2433 n = cast;
2434 }
2435
2436 return n;
2437 }
2438
2439 /**
2440 * Record profiling data from receiver profiling at an invoke with the
2441 * type system so that it can propagate it (speculation)
2442 *
2443 * @param n receiver node
2444 *
2445 * @return node with improved type
2446 */
2447 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2448 if (!UseTypeSpeculation) {
2449 return n;
2450 }
2451 ciKlass* exact_kls = profile_has_unique_klass();
2452 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2453 if ((java_bc() == Bytecodes::_checkcast ||
2454 java_bc() == Bytecodes::_instanceof ||
2455 java_bc() == Bytecodes::_aastore) &&
2456 method()->method_data()->is_mature()) {
2457 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2458 if (data != nullptr) {
2459 if (java_bc() == Bytecodes::_aastore) {
2460 ciKlass* array_type = nullptr;
2461 ciKlass* element_type = nullptr;
2462 ProfilePtrKind element_ptr = ProfileMaybeNull;
2463 bool flat_array = true;
2464 bool null_free_array = true;
2465 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2466 exact_kls = element_type;
2467 ptr_kind = element_ptr;
2468 } else {
2469 if (!data->as_BitData()->null_seen()) {
2470 ptr_kind = ProfileNeverNull;
2471 } else {
2472 assert(data->is_ReceiverTypeData(), "bad profile data type");
2473 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2474 uint i = 0;
2475 for (; i < call->row_limit(); i++) {
2476 ciKlass* receiver = call->receiver(i);
2477 if (receiver != nullptr) {
2478 break;
2479 }
2480 }
2481 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2482 }
2483 }
2484 }
2485 }
2486 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2487 }
2488
2489 /**
2490 * Record profiling data from argument profiling at an invoke with the
2491 * type system so that it can propagate it (speculation)
2492 *
2493 * @param dest_method target method for the call
2494 * @param bc what invoke bytecode is this?
2495 */
2496 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2497 if (!UseTypeSpeculation) {
2498 return;
2499 }
2500 const TypeFunc* tf = TypeFunc::make(dest_method);
2501 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2502 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2503 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2504 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2505 if (is_reference_type(targ->basic_type())) {
2506 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2507 ciKlass* better_type = nullptr;
2508 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2509 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2510 }
2511 i++;
2512 }
2513 }
2514 }
2515
2516 /**
2517 * Record profiling data from parameter profiling at an invoke with
2518 * the type system so that it can propagate it (speculation)
2519 */
2520 void GraphKit::record_profiled_parameters_for_speculation() {
2521 if (!UseTypeSpeculation) {
2522 return;
2523 }
2524 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2644 // The first null ends the list.
2645 Node* parm0, Node* parm1,
2646 Node* parm2, Node* parm3,
2647 Node* parm4, Node* parm5,
2648 Node* parm6, Node* parm7) {
2649 assert(call_addr != nullptr, "must not call null targets");
2650
2651 // Slow-path call
2652 bool is_leaf = !(flags & RC_NO_LEAF);
2653 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2654 if (call_name == nullptr) {
2655 assert(!is_leaf, "must supply name for leaf");
2656 call_name = OptoRuntime::stub_name(call_addr);
2657 }
2658 CallNode* call;
2659 if (!is_leaf) {
2660 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2661 } else if (flags & RC_NO_FP) {
2662 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2663 } else if (flags & RC_VECTOR){
2664 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2665 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2666 } else {
2667 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2668 }
2669
2670 // The following is similar to set_edges_for_java_call,
2671 // except that the memory effects of the call are restricted to AliasIdxRaw.
2672
2673 // Slow path call has no side-effects, uses few values
2674 bool wide_in = !(flags & RC_NARROW_MEM);
2675 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2676
2677 Node* prev_mem = nullptr;
2678 if (wide_in) {
2679 prev_mem = set_predefined_input_for_runtime_call(call);
2680 } else {
2681 assert(!wide_out, "narrow in => narrow out");
2682 Node* narrow_mem = memory(adr_type);
2683 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2684 }
2724
2725 if (has_io) {
2726 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2727 }
2728 return call;
2729
2730 }
2731
2732 // i2b
2733 Node* GraphKit::sign_extend_byte(Node* in) {
2734 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2735 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2736 }
2737
2738 // i2s
2739 Node* GraphKit::sign_extend_short(Node* in) {
2740 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2741 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2742 }
2743
2744
2745 //------------------------------merge_memory-----------------------------------
2746 // Merge memory from one path into the current memory state.
2747 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2748 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2749 Node* old_slice = mms.force_memory();
2750 Node* new_slice = mms.memory2();
2751 if (old_slice != new_slice) {
2752 PhiNode* phi;
2753 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2754 if (mms.is_empty()) {
2755 // clone base memory Phi's inputs for this memory slice
2756 assert(old_slice == mms.base_memory(), "sanity");
2757 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2758 _gvn.set_type(phi, Type::MEMORY);
2759 for (uint i = 1; i < phi->req(); i++) {
2760 phi->init_req(i, old_slice->in(i));
2761 }
2762 } else {
2763 phi = old_slice->as_Phi(); // Phi was generated already
2764 }
3027
3028 // Now do a linear scan of the secondary super-klass array. Again, no real
3029 // performance impact (too rare) but it's gotta be done.
3030 // Since the code is rarely used, there is no penalty for moving it
3031 // out of line, and it can only improve I-cache density.
3032 // The decision to inline or out-of-line this final check is platform
3033 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3034 Node* psc = gvn.transform(
3035 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3036
3037 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3038 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3039 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3040
3041 // Return false path; set default control to true path.
3042 *ctrl = gvn.transform(r_ok_subtype);
3043 return gvn.transform(r_not_subtype);
3044 }
3045
3046 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3047 const Type* sub_t = _gvn.type(obj_or_subklass);
3048 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3049 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3050 obj_or_subklass = makecon(sub_t);
3051 }
3052 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3053 if (expand_subtype_check) {
3054 MergeMemNode* mem = merged_memory();
3055 Node* ctrl = control();
3056 Node* subklass = obj_or_subklass;
3057 if (!sub_t->isa_klassptr()) {
3058 subklass = load_object_klass(obj_or_subklass);
3059 }
3060
3061 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3062 set_control(ctrl);
3063 return n;
3064 }
3065
3066 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3067 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3068 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3069 set_control(_gvn.transform(new IfTrueNode(iff)));
3070 return _gvn.transform(new IfFalseNode(iff));
3071 }
3072
3073 // Profile-driven exact type check:
3074 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3075 float prob, Node* *casted_receiver) {
3076 assert(!klass->is_interface(), "no exact type check on interfaces");
3077 Node* fail = top();
3078 const Type* rec_t = _gvn.type(receiver);
3079 if (rec_t->is_inlinetypeptr()) {
3080 if (klass->equals(rec_t->inline_klass())) {
3081 (*casted_receiver) = receiver; // Always passes
3082 } else {
3083 (*casted_receiver) = top(); // Always fails
3084 fail = control();
3085 set_control(top());
3086 }
3087 return fail;
3088 }
3089 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3090 Node* recv_klass = load_object_klass(receiver);
3091 fail = type_check(recv_klass, tklass, prob);
3092
3093 if (!stopped()) {
3094 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3095 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3096 assert(recv_xtype->klass_is_exact(), "");
3097
3098 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3099 // Subsume downstream occurrences of receiver with a cast to
3100 // recv_xtype, since now we know what the type will be.
3101 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3102 Node* res = _gvn.transform(cast);
3103 if (recv_xtype->is_inlinetypeptr()) {
3104 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3105 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3106 }
3107 (*casted_receiver) = res;
3108 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3109 // (User must make the replace_in_map call.)
3110 }
3111 }
3112
3113 return fail;
3114 }
3115
3116 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3117 float prob) {
3118 Node* want_klass = makecon(tklass);
3119 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3120 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3121 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3122 set_control(_gvn.transform(new IfTrueNode (iff)));
3123 Node* fail = _gvn.transform(new IfFalseNode(iff));
3124 return fail;
3125 }
3126
3127 //------------------------------subtype_check_receiver-------------------------
3128 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3129 Node** casted_receiver) {
3130 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3131 Node* want_klass = makecon(tklass);
3132
3133 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3134
3135 // Ignore interface type information until interface types are properly tracked.
3136 if (!stopped() && !klass->is_interface()) {
3137 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3138 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3139 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3140 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3141 if (recv_type->is_inlinetypeptr()) {
3142 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3143 }
3144 (*casted_receiver) = cast;
3145 }
3146 }
3147
3148 return slow_ctl;
3149 }
3150
3151 //------------------------------seems_never_null-------------------------------
3152 // Use null_seen information if it is available from the profile.
3153 // If we see an unexpected null at a type check we record it and force a
3154 // recompile; the offending check will be recompiled to handle nulls.
3155 // If we see several offending BCIs, then all checks in the
3156 // method will be recompiled.
3157 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3158 speculating = !_gvn.type(obj)->speculative_maybe_null();
3159 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3160 if (UncommonNullCast // Cutout for this technique
3161 && obj != null() // And not the -Xcomp stupid case?
3162 && !too_many_traps(reason)
3163 ) {
3164 if (speculating) {
3233
3234 //------------------------maybe_cast_profiled_receiver-------------------------
3235 // If the profile has seen exactly one type, narrow to exactly that type.
3236 // Subsequent type checks will always fold up.
3237 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3238 const TypeKlassPtr* require_klass,
3239 ciKlass* spec_klass,
3240 bool safe_for_replace) {
3241 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3242
3243 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3244
3245 // Make sure we haven't already deoptimized from this tactic.
3246 if (too_many_traps_or_recompiles(reason))
3247 return nullptr;
3248
3249 // (No, this isn't a call, but it's enough like a virtual call
3250 // to use the same ciMethod accessor to get the profile info...)
3251 // If we have a speculative type use it instead of profiling (which
3252 // may not help us)
3253 ciKlass* exact_kls = spec_klass;
3254 if (exact_kls == nullptr) {
3255 if (java_bc() == Bytecodes::_aastore) {
3256 ciKlass* array_type = nullptr;
3257 ciKlass* element_type = nullptr;
3258 ProfilePtrKind element_ptr = ProfileMaybeNull;
3259 bool flat_array = true;
3260 bool null_free_array = true;
3261 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3262 exact_kls = element_type;
3263 } else {
3264 exact_kls = profile_has_unique_klass();
3265 }
3266 }
3267 if (exact_kls != nullptr) {// no cast failures here
3268 if (require_klass == nullptr ||
3269 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3270 // If we narrow the type to match what the type profile sees or
3271 // the speculative type, we can then remove the rest of the
3272 // cast.
3273 // This is a win, even if the exact_kls is very specific,
3274 // because downstream operations, such as method calls,
3275 // will often benefit from the sharper type.
3276 Node* exact_obj = not_null_obj; // will get updated in place...
3277 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3278 &exact_obj);
3279 { PreserveJVMState pjvms(this);
3280 set_control(slow_ctl);
3281 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3282 }
3283 if (safe_for_replace) {
3284 replace_in_map(not_null_obj, exact_obj);
3285 }
3286 return exact_obj;
3376 // If not_null_obj is dead, only null-path is taken
3377 if (stopped()) { // Doing instance-of on a null?
3378 set_control(null_ctl);
3379 return intcon(0);
3380 }
3381 region->init_req(_null_path, null_ctl);
3382 phi ->init_req(_null_path, intcon(0)); // Set null path value
3383 if (null_ctl == top()) {
3384 // Do this eagerly, so that pattern matches like is_diamond_phi
3385 // will work even during parsing.
3386 assert(_null_path == PATH_LIMIT-1, "delete last");
3387 region->del_req(_null_path);
3388 phi ->del_req(_null_path);
3389 }
3390
3391 // Do we know the type check always succeed?
3392 bool known_statically = false;
3393 if (_gvn.type(superklass)->singleton()) {
3394 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3395 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3396 if (subk != nullptr && subk->is_loaded()) {
3397 int static_res = C->static_subtype_check(superk, subk);
3398 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3399 }
3400 }
3401
3402 if (!known_statically) {
3403 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3404 // We may not have profiling here or it may not help us. If we
3405 // have a speculative type use it to perform an exact cast.
3406 ciKlass* spec_obj_type = obj_type->speculative_type();
3407 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3408 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3409 if (stopped()) { // Profile disagrees with this path.
3410 set_control(null_ctl); // Null is the only remaining possibility.
3411 return intcon(0);
3412 }
3413 if (cast_obj != nullptr) {
3414 not_null_obj = cast_obj;
3415 }
3416 }
3432 record_for_igvn(region);
3433
3434 // If we know the type check always succeeds then we don't use the
3435 // profiling data at this bytecode. Don't lose it, feed it to the
3436 // type system as a speculative type.
3437 if (safe_for_replace) {
3438 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3439 replace_in_map(obj, casted_obj);
3440 }
3441
3442 return _gvn.transform(phi);
3443 }
3444
3445 //-------------------------------gen_checkcast---------------------------------
3446 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3447 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3448 // uncommon-trap paths work. Adjust stack after this call.
3449 // If failure_control is supplied and not null, it is filled in with
3450 // the control edge for the cast failure. Otherwise, an appropriate
3451 // uncommon trap or exception is thrown.
3452 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3453 kill_dead_locals(); // Benefit all the uncommon traps
3454 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3455 const Type* obj_type = _gvn.type(obj);
3456 if (obj_type->is_inlinetypeptr() && !obj_type->maybe_null() && klass_ptr_type->klass_is_exact() && obj_type->inline_klass() == klass_ptr_type->exact_klass(true)) {
3457 // Special case: larval inline objects must not be scalarized. They are also generally not
3458 // allowed to participate in most operations except as the first operand of putfield, or as an
3459 // argument to a constructor invocation with it being a receiver, Unsafe::putXXX with it being
3460 // the first argument, or Unsafe::finishPrivateBuffer. This allows us to aggressively scalarize
3461 // value objects in all other places. This special case comes from the limitation of the Java
3462 // language, Unsafe::makePrivateBuffer returns an Object that is checkcast-ed to the concrete
3463 // value type. We must do this first because C->static_subtype_check may do nothing when
3464 // StressReflectiveCode is set.
3465 return obj;
3466 }
3467
3468 // Else it must be a non-larval object
3469 obj = cast_to_non_larval(obj);
3470
3471 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3472 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3473 bool safe_for_replace = (failure_control == nullptr);
3474 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3475
3476 // Fast cutout: Check the case that the cast is vacuously true.
3477 // This detects the common cases where the test will short-circuit
3478 // away completely. We do this before we perform the null check,
3479 // because if the test is going to turn into zero code, we don't
3480 // want a residual null check left around. (Causes a slowdown,
3481 // for example, in some objArray manipulations, such as a[i]=a[j].)
3482 if (improved_klass_ptr_type->singleton()) {
3483 const TypeKlassPtr* kptr = nullptr;
3484 if (obj_type->isa_oop_ptr()) {
3485 kptr = obj_type->is_oopptr()->as_klass_type();
3486 } else if (obj->is_InlineType()) {
3487 ciInlineKlass* vk = obj_type->inline_klass();
3488 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3489 }
3490
3491 if (kptr != nullptr) {
3492 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3493 case Compile::SSC_always_true:
3494 // If we know the type check always succeed then we don't use
3495 // the profiling data at this bytecode. Don't lose it, feed it
3496 // to the type system as a speculative type.
3497 obj = record_profiled_receiver_for_speculation(obj);
3498 if (null_free) {
3499 assert(safe_for_replace, "must be");
3500 obj = null_check(obj);
3501 }
3502 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3503 return obj;
3504 case Compile::SSC_always_false:
3505 if (null_free) {
3506 assert(safe_for_replace, "must be");
3507 obj = null_check(obj);
3508 }
3509 // It needs a null check because a null will *pass* the cast check.
3510 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3511 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3512 Deoptimization::DeoptReason reason = is_aastore ?
3513 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3514 builtin_throw(reason);
3515 return top();
3516 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3517 return null_assert(obj);
3518 }
3519 break; // Fall through to full check
3520 default:
3521 break;
3522 }
3523 }
3524 }
3525
3526 ciProfileData* data = nullptr;
3527 if (failure_control == nullptr) { // use MDO in regular case only
3528 assert(java_bc() == Bytecodes::_aastore ||
3529 java_bc() == Bytecodes::_checkcast,
3530 "interpreter profiles type checks only for these BCs");
3531 if (method()->method_data()->is_mature()) {
3532 data = method()->method_data()->bci_to_data(bci());
3533 }
3534 }
3535
3536 // Make the merge point
3537 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3538 RegionNode* region = new RegionNode(PATH_LIMIT);
3539 Node* phi = new PhiNode(region, toop);
3540 _gvn.set_type(region, Type::CONTROL);
3541 _gvn.set_type(phi, toop);
3542
3543 C->set_has_split_ifs(true); // Has chance for split-if optimization
3544
3545 // Use null-cast information if it is available
3546 bool speculative_not_null = false;
3547 bool never_see_null = ((failure_control == nullptr) // regular case only
3548 && seems_never_null(obj, data, speculative_not_null));
3549
3550 if (obj->is_InlineType()) {
3551 // Re-execute if buffering during triggers deoptimization
3552 PreserveReexecuteState preexecs(this);
3553 jvms()->set_should_reexecute(true);
3554 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3555 }
3556
3557 // Null check; get casted pointer; set region slot 3
3558 Node* null_ctl = top();
3559 Node* not_null_obj = nullptr;
3560 if (null_free) {
3561 assert(safe_for_replace, "must be");
3562 not_null_obj = null_check(obj);
3563 } else {
3564 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3565 }
3566
3567 // If not_null_obj is dead, only null-path is taken
3568 if (stopped()) { // Doing instance-of on a null?
3569 set_control(null_ctl);
3570 if (toop->is_inlinetypeptr()) {
3571 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3572 }
3573 return null();
3574 }
3575 region->init_req(_null_path, null_ctl);
3576 phi ->init_req(_null_path, null()); // Set null path value
3577 if (null_ctl == top()) {
3578 // Do this eagerly, so that pattern matches like is_diamond_phi
3579 // will work even during parsing.
3580 assert(_null_path == PATH_LIMIT-1, "delete last");
3581 region->del_req(_null_path);
3582 phi ->del_req(_null_path);
3583 }
3584
3585 Node* cast_obj = nullptr;
3586 if (improved_klass_ptr_type->klass_is_exact()) {
3587 // The following optimization tries to statically cast the speculative type of the object
3588 // (for example obtained during profiling) to the type of the superklass and then do a
3589 // dynamic check that the type of the object is what we expect. To work correctly
3590 // for checkcast and aastore the type of superklass should be exact.
3591 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3592 // We may not have profiling here or it may not help us. If we have
3593 // a speculative type use it to perform an exact cast.
3594 ciKlass* spec_obj_type = obj_type->speculative_type();
3595 if (spec_obj_type != nullptr || data != nullptr) {
3596 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3597 if (cast_obj != nullptr) {
3598 if (failure_control != nullptr) // failure is now impossible
3599 (*failure_control) = top();
3600 // adjust the type of the phi to the exact klass:
3601 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3602 }
3603 }
3604 }
3605
3606 if (cast_obj == nullptr) {
3607 // Generate the subtype check
3608 Node* improved_superklass = superklass;
3609 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3610 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3611 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3612 // Additionally, the benefit would only be minor in non-constant cases.
3613 improved_superklass = makecon(improved_klass_ptr_type);
3614 }
3615 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3616 // Plug in success path into the merge
3617 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3618 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3619 if (failure_control == nullptr) {
3620 if (not_subtype_ctrl != top()) { // If failure is possible
3621 PreserveJVMState pjvms(this);
3622 set_control(not_subtype_ctrl);
3623 Node* obj_klass = nullptr;
3624 if (not_null_obj->is_InlineType()) {
3625 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3626 } else {
3627 obj_klass = load_object_klass(not_null_obj);
3628 }
3629 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3630 Deoptimization::DeoptReason reason = is_aastore ?
3631 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3632 builtin_throw(reason);
3633 }
3634 } else {
3635 (*failure_control) = not_subtype_ctrl;
3636 }
3637 }
3638
3639 region->init_req(_obj_path, control());
3640 phi ->init_req(_obj_path, cast_obj);
3641
3642 // A merge of null or Casted-NotNull obj
3643 Node* res = _gvn.transform(phi);
3644
3645 // Note I do NOT always 'replace_in_map(obj,result)' here.
3646 // if( tk->klass()->can_be_primary_super() )
3647 // This means that if I successfully store an Object into an array-of-String
3648 // I 'forget' that the Object is really now known to be a String. I have to
3649 // do this because we don't have true union types for interfaces - if I store
3650 // a Baz into an array-of-Interface and then tell the optimizer it's an
3651 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3652 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3653 // replace_in_map( obj, res );
3654
3655 // Return final merged results
3656 set_control( _gvn.transform(region) );
3657 record_for_igvn(region);
3658
3659 bool not_inline = !toop->can_be_inline_type();
3660 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3661 if (EnableValhalla && (not_inline || not_flat_in_array)) {
3662 // Check if obj has been loaded from an array
3663 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3664 Node* array = nullptr;
3665 if (obj->isa_Load()) {
3666 Node* address = obj->in(MemNode::Address);
3667 if (address->isa_AddP()) {
3668 array = address->as_AddP()->in(AddPNode::Base);
3669 }
3670 } else if (obj->is_Phi()) {
3671 Node* region = obj->in(0);
3672 // TODO make this more robust (see JDK-8231346)
3673 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3674 IfNode* iff = region->in(2)->in(0)->isa_If();
3675 if (iff != nullptr) {
3676 iff->is_flat_array_check(&_gvn, &array);
3677 }
3678 }
3679 }
3680 if (array != nullptr) {
3681 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3682 if (ary_t != nullptr) {
3683 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3684 // Casting array element to a non-inline-type, mark array as not null-free.
3685 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3686 replace_in_map(array, cast);
3687 array = cast;
3688 }
3689 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3690 // Casting array element to a non-flat-in-array type, mark array as not flat.
3691 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3692 replace_in_map(array, cast);
3693 array = cast;
3694 }
3695 }
3696 }
3697 }
3698
3699 if (!stopped() && !res->is_InlineType()) {
3700 res = record_profiled_receiver_for_speculation(res);
3701 if (toop->is_inlinetypeptr() && !maybe_larval) {
3702 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3703 res = vt;
3704 if (safe_for_replace) {
3705 replace_in_map(obj, vt);
3706 replace_in_map(not_null_obj, vt);
3707 replace_in_map(res, vt);
3708 }
3709 }
3710 }
3711 return res;
3712 }
3713
3714 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3715 // Load markword
3716 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3717 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3718 if (check_lock) {
3719 // Check if obj is locked
3720 Node* locked_bit = MakeConX(markWord::unlocked_value);
3721 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3722 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3723 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3724 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3725 _gvn.transform(iff);
3726 Node* locked_region = new RegionNode(3);
3727 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3728
3729 // Unlocked: Use bits from mark word
3730 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3731 mark_phi->init_req(1, mark);
3732
3733 // Locked: Load prototype header from klass
3734 set_control(_gvn.transform(new IfFalseNode(iff)));
3735 // Make loads control dependent to make sure they are only executed if array is locked
3736 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3737 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3738 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3739 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3740
3741 locked_region->init_req(2, control());
3742 mark_phi->init_req(2, proto);
3743 set_control(_gvn.transform(locked_region));
3744 record_for_igvn(locked_region);
3745
3746 mark = mark_phi;
3747 }
3748
3749 // Now check if mark word bits are set
3750 Node* mask = MakeConX(mask_val);
3751 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3752 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3753 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3754 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3755 }
3756
3757 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3758 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3759 }
3760
3761 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3762 // We can't use immutable memory here because the mark word is mutable.
3763 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3764 // check is moved out of loops (mainly to enable loop unswitching).
3765 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3766 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3767 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3768 }
3769
3770 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3771 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3772 }
3773
3774 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3775 assert(vk->has_atomic_layout() || vk->has_non_atomic_layout(), "Can't be null-free and flat");
3776
3777 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3778 if (!vk->has_non_atomic_layout()) {
3779 return intcon(1); // Always atomic
3780 } else if (!vk->has_atomic_layout()) {
3781 return intcon(0); // Never atomic
3782 }
3783
3784 // TODO 8350865 Don't fold this klass load because atomicity is currently not included in the typesystem
3785 Node* array_klass = load_object_klass(array, /* fold_for_arrays = */ false);
3786 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3787 Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset);
3788 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3789 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::ATOMIC_FLAT)));
3790 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3791 }
3792
3793 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3794 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3795 RegionNode* region = new RegionNode(3);
3796 Node* null_ctl = top();
3797 null_check_oop(val, &null_ctl);
3798 if (null_ctl != top()) {
3799 PreserveJVMState pjvms(this);
3800 set_control(null_ctl);
3801 {
3802 // Deoptimize if null-free array
3803 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3804 inc_sp(nargs);
3805 uncommon_trap(Deoptimization::Reason_null_check,
3806 Deoptimization::Action_none);
3807 }
3808 region->init_req(1, control());
3809 }
3810 region->init_req(2, control());
3811 set_control(_gvn.transform(region));
3812 record_for_igvn(region);
3813 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3814 // Since we were just successfully storing null, the array can't be null free.
3815 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3816 ary_t = ary_t->cast_to_not_null_free();
3817 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3818 if (safe_for_replace) {
3819 replace_in_map(ary, cast);
3820 }
3821 ary = cast;
3822 }
3823 return ary;
3824 }
3825
3826 //------------------------------next_monitor-----------------------------------
3827 // What number should be given to the next monitor?
3828 int GraphKit::next_monitor() {
3829 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3830 int next = current + C->sync_stack_slots();
3831 // Keep the toplevel high water mark current:
3832 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3833 return current;
3834 }
3835
3836 //------------------------------insert_mem_bar---------------------------------
3837 // Memory barrier to avoid floating things around
3838 // The membar serves as a pinch point between both control and all memory slices.
3839 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3840 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3841 mb->init_req(TypeFunc::Control, control());
3842 mb->init_req(TypeFunc::Memory, reset_memory());
3843 Node* membar = _gvn.transform(mb);
3871 }
3872 Node* membar = _gvn.transform(mb);
3873 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3874 if (alias_idx == Compile::AliasIdxBot) {
3875 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3876 } else {
3877 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3878 }
3879 return membar;
3880 }
3881
3882 //------------------------------shared_lock------------------------------------
3883 // Emit locking code.
3884 FastLockNode* GraphKit::shared_lock(Node* obj) {
3885 // bci is either a monitorenter bc or InvocationEntryBci
3886 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3887 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3888
3889 if( !GenerateSynchronizationCode )
3890 return nullptr; // Not locking things?
3891
3892 if (stopped()) // Dead monitor?
3893 return nullptr;
3894
3895 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3896
3897 // Box the stack location
3898 Node* box = new BoxLockNode(next_monitor());
3899 // Check for bailout after new BoxLockNode
3900 if (failing()) { return nullptr; }
3901 box = _gvn.transform(box);
3902 Node* mem = reset_memory();
3903
3904 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3905
3906 // Add monitor to debug info for the slow path. If we block inside the
3907 // slow path and de-opt, we need the monitor hanging around
3908 map()->push_monitor( flock );
3909
3910 const TypeFunc *tf = LockNode::lock_type();
3911 LockNode *lock = new LockNode(C, tf);
3940 }
3941 #endif
3942
3943 return flock;
3944 }
3945
3946
3947 //------------------------------shared_unlock----------------------------------
3948 // Emit unlocking code.
3949 void GraphKit::shared_unlock(Node* box, Node* obj) {
3950 // bci is either a monitorenter bc or InvocationEntryBci
3951 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3952 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3953
3954 if( !GenerateSynchronizationCode )
3955 return;
3956 if (stopped()) { // Dead monitor?
3957 map()->pop_monitor(); // Kill monitor from debug info
3958 return;
3959 }
3960 assert(!obj->is_InlineType(), "should not unlock on inline type");
3961
3962 // Memory barrier to avoid floating things down past the locked region
3963 insert_mem_bar(Op_MemBarReleaseLock);
3964
3965 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3966 UnlockNode *unlock = new UnlockNode(C, tf);
3967 #ifdef ASSERT
3968 unlock->set_dbg_jvms(sync_jvms());
3969 #endif
3970 uint raw_idx = Compile::AliasIdxRaw;
3971 unlock->init_req( TypeFunc::Control, control() );
3972 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3973 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3974 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3975 unlock->init_req( TypeFunc::ReturnAdr, top() );
3976
3977 unlock->init_req(TypeFunc::Parms + 0, obj);
3978 unlock->init_req(TypeFunc::Parms + 1, box);
3979 unlock = _gvn.transform(unlock)->as_Unlock();
3980
3981 Node* mem = reset_memory();
3982
3983 // unlock has no side-effects, sets few values
3984 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3985
3986 // Kill monitor from debug info
3987 map()->pop_monitor( );
3988 }
3989
3990 //-------------------------------get_layout_helper-----------------------------
3991 // If the given klass is a constant or known to be an array,
3992 // fetch the constant layout helper value into constant_value
3993 // and return null. Otherwise, load the non-constant
3994 // layout helper value, and return the node which represents it.
3995 // This two-faced routine is useful because allocation sites
3996 // almost always feature constant types.
3997 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3998 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3999 if (!StressReflectiveCode && klass_t != nullptr) {
4000 bool xklass = klass_t->klass_is_exact();
4001 bool can_be_flat = false;
4002 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4003 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4004 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4005 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4006 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4007 }
4008 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4009 jint lhelper;
4010 if (klass_t->is_flat()) {
4011 lhelper = ary_type->flat_layout_helper();
4012 } else if (klass_t->isa_aryklassptr()) {
4013 BasicType elem = ary_type->elem()->array_element_basic_type();
4014 if (is_reference_type(elem, true)) {
4015 elem = T_OBJECT;
4016 }
4017 lhelper = Klass::array_layout_helper(elem);
4018 } else {
4019 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4020 }
4021 if (lhelper != Klass::_lh_neutral_value) {
4022 constant_value = lhelper;
4023 return (Node*) nullptr;
4024 }
4025 }
4026 }
4027 constant_value = Klass::_lh_neutral_value; // put in a known value
4028 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4029 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4030 }
4031
4032 // We just put in an allocate/initialize with a big raw-memory effect.
4033 // Hook selected additional alias categories on the initialization.
4034 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4035 MergeMemNode* init_in_merge,
4036 Node* init_out_raw) {
4037 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4038 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4039
4040 Node* prevmem = kit.memory(alias_idx);
4041 init_in_merge->set_memory_at(alias_idx, prevmem);
4042 if (init_out_raw != nullptr) {
4043 kit.set_memory(init_out_raw, alias_idx);
4044 }
4045 }
4046
4047 //---------------------------set_output_for_allocation-------------------------
4048 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4049 const TypeOopPtr* oop_type,
4050 bool deoptimize_on_exception) {
4051 int rawidx = Compile::AliasIdxRaw;
4052 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4053 add_safepoint_edges(alloc);
4054 Node* allocx = _gvn.transform(alloc);
4055 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4056 // create memory projection for i_o
4057 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4058 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4059
4060 // create a memory projection as for the normal control path
4061 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4062 set_memory(malloc, rawidx);
4063
4064 // a normal slow-call doesn't change i_o, but an allocation does
4065 // we create a separate i_o projection for the normal control path
4066 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4067 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4068
4069 // put in an initialization barrier
4070 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4071 rawoop)->as_Initialize();
4072 assert(alloc->initialization() == init, "2-way macro link must work");
4073 assert(init ->allocation() == alloc, "2-way macro link must work");
4074 {
4075 // Extract memory strands which may participate in the new object's
4076 // initialization, and source them from the new InitializeNode.
4077 // This will allow us to observe initializations when they occur,
4078 // and link them properly (as a group) to the InitializeNode.
4079 assert(init->in(InitializeNode::Memory) == malloc, "");
4080 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4081 init->set_req(InitializeNode::Memory, minit_in);
4082 record_for_igvn(minit_in); // fold it up later, if possible
4083 _gvn.set_type(minit_in, Type::MEMORY);
4084 Node* minit_out = memory(rawidx);
4085 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4086 // Add an edge in the MergeMem for the header fields so an access
4087 // to one of those has correct memory state
4088 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4089 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4090 if (oop_type->isa_aryptr()) {
4091 const TypeAryPtr* arytype = oop_type->is_aryptr();
4092 if (arytype->is_flat()) {
4093 // Initially all flat array accesses share a single slice
4094 // but that changes after parsing. Prepare the memory graph so
4095 // it can optimize flat array accesses properly once they
4096 // don't share a single slice.
4097 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4098 C->set_flat_accesses_share_alias(false);
4099 ciInlineKlass* vk = arytype->elem()->inline_klass();
4100 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4101 ciField* field = vk->nonstatic_field_at(i);
4102 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4103 continue; // do not bother to track really large numbers of fields
4104 int off_in_vt = field->offset_in_bytes() - vk->payload_offset();
4105 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4106 int fieldidx = C->get_alias_index(adr_type, true);
4107 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
4108 // can result in per flat array field Phis to be created which confuses the logic of
4109 // Compile::adjust_flat_array_access_aliases().
4110 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
4111 }
4112 C->set_flat_accesses_share_alias(true);
4113 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4114 } else {
4115 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4116 int elemidx = C->get_alias_index(telemref);
4117 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4118 }
4119 } else if (oop_type->isa_instptr()) {
4120 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4121 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4122 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4123 ciField* field = ik->nonstatic_field_at(i);
4124 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4125 continue; // do not bother to track really large numbers of fields
4126 // Find (or create) the alias category for this field:
4127 int fieldidx = C->alias_type(field)->index();
4128 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4129 }
4130 }
4131 }
4132
4133 // Cast raw oop to the real thing...
4134 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4135 javaoop = _gvn.transform(javaoop);
4136 C->set_recent_alloc(control(), javaoop);
4137 assert(just_allocated_object(control()) == javaoop, "just allocated");
4138
4139 #ifdef ASSERT
4140 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4151 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4152 }
4153 }
4154 #endif //ASSERT
4155
4156 return javaoop;
4157 }
4158
4159 //---------------------------new_instance--------------------------------------
4160 // This routine takes a klass_node which may be constant (for a static type)
4161 // or may be non-constant (for reflective code). It will work equally well
4162 // for either, and the graph will fold nicely if the optimizer later reduces
4163 // the type to a constant.
4164 // The optional arguments are for specialized use by intrinsics:
4165 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4166 // - If 'return_size_val', report the total object size to the caller.
4167 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4168 Node* GraphKit::new_instance(Node* klass_node,
4169 Node* extra_slow_test,
4170 Node* *return_size_val,
4171 bool deoptimize_on_exception,
4172 InlineTypeNode* inline_type_node) {
4173 // Compute size in doublewords
4174 // The size is always an integral number of doublewords, represented
4175 // as a positive bytewise size stored in the klass's layout_helper.
4176 // The layout_helper also encodes (in a low bit) the need for a slow path.
4177 jint layout_con = Klass::_lh_neutral_value;
4178 Node* layout_val = get_layout_helper(klass_node, layout_con);
4179 bool layout_is_con = (layout_val == nullptr);
4180
4181 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4182 // Generate the initial go-slow test. It's either ALWAYS (return a
4183 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4184 // case) a computed value derived from the layout_helper.
4185 Node* initial_slow_test = nullptr;
4186 if (layout_is_con) {
4187 assert(!StressReflectiveCode, "stress mode does not use these paths");
4188 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4189 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4190 } else { // reflective case
4191 // This reflective path is used by Unsafe.allocateInstance.
4192 // (It may be stress-tested by specifying StressReflectiveCode.)
4193 // Basically, we want to get into the VM is there's an illegal argument.
4194 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4195 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4196 if (extra_slow_test != intcon(0)) {
4197 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4198 }
4199 // (Macro-expander will further convert this to a Bool, if necessary.)
4210
4211 // Clear the low bits to extract layout_helper_size_in_bytes:
4212 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4213 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4214 size = _gvn.transform( new AndXNode(size, mask) );
4215 }
4216 if (return_size_val != nullptr) {
4217 (*return_size_val) = size;
4218 }
4219
4220 // This is a precise notnull oop of the klass.
4221 // (Actually, it need not be precise if this is a reflective allocation.)
4222 // It's what we cast the result to.
4223 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4224 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4225 const TypeOopPtr* oop_type = tklass->as_instance_type();
4226
4227 // Now generate allocation code
4228
4229 // The entire memory state is needed for slow path of the allocation
4230 // since GC and deoptimization can happen.
4231 Node *mem = reset_memory();
4232 set_all_memory(mem); // Create new memory state
4233
4234 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4235 control(), mem, i_o(),
4236 size, klass_node,
4237 initial_slow_test, inline_type_node);
4238
4239 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4240 }
4241
4242 //-------------------------------new_array-------------------------------------
4243 // helper for newarray and anewarray
4244 // The 'length' parameter is (obviously) the length of the array.
4245 // The optional arguments are for specialized use by intrinsics:
4246 // - If 'return_size_val', report the non-padded array size (sum of header size
4247 // and array body) to the caller.
4248 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4249 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4250 Node* length, // number of array elements
4251 int nargs, // number of arguments to push back for uncommon trap
4252 Node* *return_size_val,
4253 bool deoptimize_on_exception,
4254 Node* init_val) {
4255 jint layout_con = Klass::_lh_neutral_value;
4256 Node* layout_val = get_layout_helper(klass_node, layout_con);
4257 bool layout_is_con = (layout_val == nullptr);
4258
4259 if (!layout_is_con && !StressReflectiveCode &&
4260 !too_many_traps(Deoptimization::Reason_class_check)) {
4261 // This is a reflective array creation site.
4262 // Optimistically assume that it is a subtype of Object[],
4263 // so that we can fold up all the address arithmetic.
4264 layout_con = Klass::array_layout_helper(T_OBJECT);
4265 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4266 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4267 { BuildCutout unless(this, bol_lh, PROB_MAX);
4268 inc_sp(nargs);
4269 uncommon_trap(Deoptimization::Reason_class_check,
4270 Deoptimization::Action_maybe_recompile);
4271 }
4272 layout_val = nullptr;
4273 layout_is_con = true;
4274 }
4275
4276 // Generate the initial go-slow test. Make sure we do not overflow
4277 // if length is huge (near 2Gig) or negative! We do not need
4278 // exact double-words here, just a close approximation of needed
4279 // double-words. We can't add any offset or rounding bits, lest we
4280 // take a size -1 of bytes and make it positive. Use an unsigned
4281 // compare, so negative sizes look hugely positive.
4282 int fast_size_limit = FastAllocateSizeLimit;
4283 if (layout_is_con) {
4284 assert(!StressReflectiveCode, "stress mode does not use these paths");
4285 // Increase the size limit if we have exact knowledge of array type.
4286 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4287 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4288 }
4289
4290 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4291 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4292
4293 // --- Size Computation ---
4294 // array_size = round_to_heap(array_header + (length << elem_shift));
4295 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4296 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4297 // The rounding mask is strength-reduced, if possible.
4298 int round_mask = MinObjAlignmentInBytes - 1;
4299 Node* header_size = nullptr;
4300 // (T_BYTE has the weakest alignment and size restrictions...)
4301 if (layout_is_con) {
4302 int hsize = Klass::layout_helper_header_size(layout_con);
4303 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4304 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4305 if ((round_mask & ~right_n_bits(eshift)) == 0)
4306 round_mask = 0; // strength-reduce it if it goes away completely
4307 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4308 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4309 assert(header_size_min <= hsize, "generic minimum is smallest");
4310 header_size = intcon(hsize);
4311 } else {
4312 Node* hss = intcon(Klass::_lh_header_size_shift);
4313 Node* hsm = intcon(Klass::_lh_header_size_mask);
4314 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4315 header_size = _gvn.transform(new AndINode(header_size, hsm));
4316 }
4317
4318 Node* elem_shift = nullptr;
4319 if (layout_is_con) {
4320 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4321 if (eshift != 0)
4322 elem_shift = intcon(eshift);
4323 } else {
4324 // There is no need to mask or shift this value.
4325 // The semantics of LShiftINode include an implicit mask to 0x1F.
4326 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4327 elem_shift = layout_val;
4374 }
4375 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4376
4377 if (return_size_val != nullptr) {
4378 // This is the size
4379 (*return_size_val) = non_rounded_size;
4380 }
4381
4382 Node* size = non_rounded_size;
4383 if (round_mask != 0) {
4384 Node* mask1 = MakeConX(round_mask);
4385 size = _gvn.transform(new AddXNode(size, mask1));
4386 Node* mask2 = MakeConX(~round_mask);
4387 size = _gvn.transform(new AndXNode(size, mask2));
4388 }
4389 // else if round_mask == 0, the size computation is self-rounding
4390
4391 // Now generate allocation code
4392
4393 // The entire memory state is needed for slow path of the allocation
4394 // since GC and deoptimization can happen.
4395 Node *mem = reset_memory();
4396 set_all_memory(mem); // Create new memory state
4397
4398 if (initial_slow_test->is_Bool()) {
4399 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4400 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4401 }
4402
4403 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4404 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4405
4406 Node* raw_init_value = nullptr;
4407 if (init_val != nullptr) {
4408 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4409 if (ary_type->is_flat()) {
4410 initial_slow_test = intcon(1);
4411 }
4412
4413 if (UseCompressedOops) {
4414 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4415 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4416 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4417 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4418 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4419 } else {
4420 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4421 }
4422 }
4423
4424 Node* valid_length_test = _gvn.intcon(1);
4425 if (ary_type->isa_aryptr()) {
4426 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4427 jint max = TypeAryPtr::max_array_length(bt);
4428 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4429 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4430 }
4431
4432 // Create the AllocateArrayNode and its result projections
4433 AllocateArrayNode* alloc
4434 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4435 control(), mem, i_o(),
4436 size, klass_node,
4437 initial_slow_test,
4438 length, valid_length_test,
4439 init_val, raw_init_value);
4440 // Cast to correct type. Note that the klass_node may be constant or not,
4441 // and in the latter case the actual array type will be inexact also.
4442 // (This happens via a non-constant argument to inline_native_newArray.)
4443 // In any case, the value of klass_node provides the desired array type.
4444 const TypeInt* length_type = _gvn.find_int_type(length);
4445 if (ary_type->isa_aryptr() && length_type != nullptr) {
4446 // Try to get a better type than POS for the size
4447 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4448 }
4449
4450 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4451
4452 array_ideal_length(alloc, ary_type, true);
4453 return javaoop;
4454 }
4455
4456 // The following "Ideal_foo" functions are placed here because they recognize
4457 // the graph shapes created by the functions immediately above.
4458
4459 //---------------------------Ideal_allocation----------------------------------
4567 set_all_memory(ideal.merged_memory());
4568 set_i_o(ideal.i_o());
4569 set_control(ideal.ctrl());
4570 }
4571
4572 void GraphKit::final_sync(IdealKit& ideal) {
4573 // Final sync IdealKit and graphKit.
4574 sync_kit(ideal);
4575 }
4576
4577 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4578 Node* len = load_array_length(load_String_value(str, set_ctrl));
4579 Node* coder = load_String_coder(str, set_ctrl);
4580 // Divide length by 2 if coder is UTF16
4581 return _gvn.transform(new RShiftINode(len, coder));
4582 }
4583
4584 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4585 int value_offset = java_lang_String::value_offset();
4586 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4587 false, nullptr, Type::Offset(0));
4588 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4589 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4590 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4591 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4592 Node* p = basic_plus_adr(str, str, value_offset);
4593 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4594 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4595 return load;
4596 }
4597
4598 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4599 if (!CompactStrings) {
4600 return intcon(java_lang_String::CODER_UTF16);
4601 }
4602 int coder_offset = java_lang_String::coder_offset();
4603 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4604 false, nullptr, Type::Offset(0));
4605 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4606
4607 Node* p = basic_plus_adr(str, str, coder_offset);
4608 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4609 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4610 return load;
4611 }
4612
4613 void GraphKit::store_String_value(Node* str, Node* value) {
4614 int value_offset = java_lang_String::value_offset();
4615 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4616 false, nullptr, Type::Offset(0));
4617 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4618
4619 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4620 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4621 }
4622
4623 void GraphKit::store_String_coder(Node* str, Node* value) {
4624 int coder_offset = java_lang_String::coder_offset();
4625 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4626 false, nullptr, Type::Offset(0));
4627 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4628
4629 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4630 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4631 }
4632
4633 // Capture src and dst memory state with a MergeMemNode
4634 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4635 if (src_type == dst_type) {
4636 // Types are equal, we don't need a MergeMemNode
4637 return memory(src_type);
4638 }
4639 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4640 record_for_igvn(merge); // fold it up later, if possible
4641 int src_idx = C->get_alias_index(src_type);
4642 int dst_idx = C->get_alias_index(dst_type);
4643 merge->set_memory_at(src_idx, memory(src_idx));
4644 merge->set_memory_at(dst_idx, memory(dst_idx));
4645 return merge;
4646 }
4719 i_char->init_req(2, AddI(i_char, intcon(2)));
4720
4721 set_control(IfFalse(iff));
4722 set_memory(st, TypeAryPtr::BYTES);
4723 }
4724
4725 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4726 if (!field->is_constant()) {
4727 return nullptr; // Field not marked as constant.
4728 }
4729 ciInstance* holder = nullptr;
4730 if (!field->is_static()) {
4731 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4732 if (const_oop != nullptr && const_oop->is_instance()) {
4733 holder = const_oop->as_instance();
4734 }
4735 }
4736 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4737 /*is_unsigned_load=*/false);
4738 if (con_type != nullptr) {
4739 Node* con = makecon(con_type);
4740 if (field->type()->is_inlinetype()) {
4741 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4742 } else if (con_type->is_inlinetypeptr()) {
4743 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4744 }
4745 return con;
4746 }
4747 return nullptr;
4748 }
4749
4750 //---------------------------load_mirror_from_klass----------------------------
4751 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4752 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4753 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4754 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4755 // mirror = ((OopHandle)mirror)->resolve();
4756 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4757 }
4758
4759 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4760 const Type* obj_type = obj->bottom_type();
4761 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4762 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4763 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4764 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4765 obj = casted_obj;
4766 }
4767 if (sig_type->is_inlinetypeptr()) {
4768 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4769 }
4770 return obj;
4771 }
|